https://rodneybrooks.com/predictions-scorecard-2025-january-01/ Rodney Brooks Robots, AI, and other stuff * Blog * MIT * Robust.AI [robots] Search for: [ ] [Search] Recent Posts * Predictions Scorecard, 2025 January 01 * Tips For Building and Deploying Robots * Rodney Brooks' Three Laws of Artificial Intelligence * Rodney Brooks' Three Laws of Robotics * Predictions Scorecard, 2024 January 01 Categories * Dated Predictions * Essays * Quick Takes * Reviews * Whimsy Archives * January 2025 * September 2024 * July 2024 * January 2024 * December 2023 * October 2023 * March 2023 * January 2023 * November 2022 * May 2022 * January 2022 * February 2021 * January 2021 * December 2020 * August 2020 * May 2020 * April 2020 * January 2020 * May 2019 * March 2019 * January 2019 * July 2018 * April 2018 * March 2018 * February 2018 * January 2018 * December 2017 * September 2017 * August 2017 * June 2017 * May 2017 * April 2017 * March 2017 * February 2017 * January 2017 Subscribe Receive email notifications: Leave This Blank:[ ]Leave This Blank Too: [ ]Do Not Change This:[http:// ] Your email: [Enter email address.] [Subscribe] [Unsubscribe] Post: Predictions Scorecard, 2025 January 01 January 1, 2025 -- Dated Predictions Predictions Scorecard, 2025 January 01 [fb] rodneybrooks.com/predictions-scorecard-2025-january-01/ [You can follow me on social media: @rodneyabrooks.bsky.social and see my publications etc., at https://people.csail.mit.edu/brooks] This is my seventh annual update on how my dated predictions from January 1^st, 2018 concerning (1) self driving cars, (2) robotics, AI , and machine learning, and (3) human space travel, have held up. I promised then to review them at the start of the year every year until 2050 (right after my 95^th birthday), thirty two years in total. The idea is to hold myself accountable for those predictions. How right or wrong was I? I have decided to change my rules for myself a little bit after this year, in response to the many many people who have said how much they enjoy seeing my updates. My predictions were mostly for the first few years, and by next year the density of due dates will be very low. So, on the eight anniversary of my first set of predictions, i.e., a year from today, I will be making a new set of predictions centered on the period January 1^st 2026 to January 1^st 2036, and that will give a new density of predictions where there will be real meat to see how accurately they turned out. What I Want to Achieve and a Changing Hype-driven Landscape The level of hype about AI, Machine Learning and Robotics completely distorts people's understanding of reality. It distorts where VC money goes, always to something that promises impossibly large payoffs-it seems it is better to have an untested idea that would have an enormous payoff than a tested idea which can get to a sustainable business, but does not change the world for ever. It distorts what young researchers work on as they do not want to be seen as old fashioned even when the current hyped topic is sort of dumb-soon the dumbness is forgotten and the heat of the chase becomes all. It distorts what people think they need to get a degree in at college in order to have good career prospects. I want people to use rational thought processes when they hear about hyped ideas and be able to assess what is really going on, and what is just plain (to use the technical term) bullshit. My Color Scheme and Past Analysis The acronyms I used for predictions in my original post were as follows. NET year means it will not happen before that year (No Earlier Than) BY year means I predict that it will happen by that year. NIML, Not In My Lifetime, i.e., not before 2050. As time passes mentioned years I color then as accurate, too pessimistic, or too optimistic. This year I have added hemming and hawing. This is for when something looks just like what I said would take a lot longer has happened, but the underlying achievement is not what everyone expected, and is not what was delivered. This is mostly for things that were talked about as being likely to happen with no human intervention and it now appears to happen that way, but in reality there are humans in the loop that the companies never disclose. So the technology that was promised to be delivered hasn't actually been delivered but everyone thinks it has been. I have not changed any of the text of the first three columns of the prediction tables since their publication on the first day of 2018. I only change the text in the fourth column to say what actually happened. This meant that by two years ago that fourth column was getting very long and skinny, so I removed them and started with fresh comments last year. I have kept last year's comments and added new ones, with yellow backgrounds, for this year. If you want to see the previous five years of comments you can go back to the 2023 scorecard. Overview of changes this year There has been a lot of activity in both self driving cars (the demise of Cruise a big push by Waymo to scale human assisted deployments, and lots of smoke and mirrors from an electric car company) and in AI, where robotics has been pulled into the ultra hyposphere while in generative AI the end of scaling and the introduction of inference mechanisms (!!) have been hotly announced and disputed. The human spaceflight endeavor, as it did last year, has crawled along and again has stretched out dates that were probably too optimistic in the first place. But First. We all know about FOMO, Fear Of Missing Out. In late 2023, for a talk on generative AI that I gave at MIT, I coined another acronym, FOBAWTPALSL, Fear Of Being A Wimpy Techno-Pessimist And Looking Stupid Later. Perhaps that one is a little bit too much of a mouthful to catch on. These two human insecurities lead people to herd-like behavior in establishing and propagating the zeitgeist on almost any topic. They lead to people piling on the hype fiestas, rushing to invest (money, effort, or hope) in marginal ideas once they have become a little bit popular, or believing our airspace is being invaded by foreign drones. "Mounting evidence, and lack thereof, suggests that perhaps the whole craze has been a sort of communal fever dream fueled by crowd mentality, confirmation bias and a general distrust in all things official." That quote is from the drone story linked to above, but it could well as been about the hype that we are moving towards AGI (Artificial General Intelligence). I want to be clear, as there has been for almost seventy years now, there has been significant progress in Artificial Intelligence over the last decade. There are new tools and they are being applied widely in science and technology, and are changing the way we think about ourselves, and how to make further progress. That being said, we are not on the verge of replacing and eliminating humans in either white collar jobs or blue collar jobs. Their tasks may shift in both styles of jobs, but the jobs are not going away. We are not on the verge of a revolution in medicine and the role of human doctors. We are not on the verge of the elimination of coding as a job. We are not on the verge of replacing humans with humanoid robots to do jobs that involve physical interactions in the world. We are not on the verge of replacing human automobile and truck drivers world wide. We are not on the verge of replacing scientists with AI programs. Breathless predictions such as these have happened for seven decades in a row, and each time people have thought the end is in sight and that it is all over for humans, that we have figured out the secrets of intelligence and it will all just scale. The only difference this time is that these expectations have leaked out into the world at large. I'll analyze why this continues to happen below in the section on AI and ML. Here is a list of some of those hype cycles that I, personally, have perceived and lived through, as taken from my presentation at MIT in late 2023 that I referenced above re FOBAWTPALSL. [AIhypecycles] Really, was there really hype about all these things? Yes, there was, within the circles that cared. Those circles have gotten wider and wider and when reigning world chess champion Garry Kasparov was beaten by I.B.M.'s Deep Blue computer under tournament conditions in 1997 it was widely reported in the popular press, And it was declared that it was all over for humans. Back in February 2011 a computer program named Watson played on the television game show Jeopardy against all time human champions. John Markoff, legendary technology reporter at the New York Times, wrote stories about this the day before the competition, and the day after, when Watson had indeed beaten the humans, with the same questions (fed as text to it as the same time as the humans heard the questions) all running on a cluster of machines not connected to an outside network. Here are three successive paragraphs from the second of those stories. For I.B.M., the future will happen very quickly, company executives said. On Thursday it plans to announce that it will collaborate with Columbia University and the University of Maryland to create a physician's assistant service that will allow doctors to query a cybernetic assistant. The company also plans to work with Nuance Communications Inc. to add voice recognition to the physician's assistant, possibly making the service available in as little as 18 months. "I have been in medical education for 40 years and we're still a very memory-based curriculum," said Dr. Herbert Chase, a professor of clinical medicine at Columbia University who is working with I.B.M. on the physician's assistant. "The power of Watson- like tools will cause us to reconsider what it is we want students to do." I.B.M. executives also said they are in discussions with a major consumer electronics retailer to develop a version of Watson, named after I.B.M.'s founder, Thomas J. Watson, that would be able to interact with consumers on a variety of subjects like buying decisions and technical support. My personal experience at that time was people I did not know, but who had heard about my role at MIT (as director of the MIT AI Lab, and then founding director of MIT CSAIL, the Computer Science and Artificial Intelligence Lab) would come up to me and ask about the future of medicine. The people were variously doctors or health industry executives. I reassured them that medicine as we knew it then would stay much the same and was not about to be rendered obsolete. And then in 2016 Geoff Hinton, one of the key architects of Deep Learning (which has had undeniable impact on the world) said: "People should stop training radiologists now. It is just completely obvious that within five years deep learning is going to be better than radiologists." More people asking me whether this was true. It wasn't in five years and it isn't now. We need more radiologists than ever. And yes they do use deep learning tools to help them see some things they wouldn't otherwise see. But they also understand anomalies using causal reasoning and we would be in a sorry state if all radiology was done by programs today. Now look at those plum colored paragraphs above again as you take yourself way back in time to a year or so ago when ChatGPT was just a baby AGI, You can find stories just like this one if you substitute "ChatGPT" for "Watson" and "Microsoft" for "I.B.M." The things confidently predicted in 2011 (and in 1979, and in 2016) about the end of doctors didn't happen then and it is not happening now. Nor are all the other jobs ending. Today I get asked about humanoid robots taking away people's jobs. In March 2023 I was at a cocktail party and there was a humanoid robot behind the bar making jokes with people and shakily (in a bad way) mixing drinks. A waiter was standing about 20 feet away silently staring at the robot with mouth hanging open. I went over and told her it was tele-operated. "Thank God" she said. (And I didn't need to explain what "tele-operated" meant). Humanoids are not going to be taking away jobs anytime soon (and by that I mean not for decades). You, you people!, are all making fundamental errors in understanding the technologies and where their boundaries lie. Many of them will be useful technologies but their imagined capabilities are just not going to come about in the time frames the majority of the technology and prognosticator class, deeply driven by FOBAWTPALSL, think. But this time it is different you say. This time it is really going to happen. You just don't understand how powerful AI is now, you say. All the early predictions were clearly wrong and premature as the AI programs were clearly not as good as now and we had much less computation back then. This time it is all different and it is for sure now. Yeah, well, I've got a Second Coming to sell you... Self Driving Cars As with flying cars the definition, or common understanding, of what self driving cars really means has changed since my post on predictions seven years ago. At that time self driving cars meant that the cars would drive themselves to wherever they were told to go with no further human control inputs. Now self driving cars means that there is no one in the drivers seat, but there may well be, and in all cases so far deployed, humans monitoring those cars from a remote location, and occasionally sending control inputs to the cars. The companies do not advertise this feature out loud too much, but they do acknowledge it, and the reports are that it happens somewhere between every one to two miles traveled. These inputs are not direct control of the normal human mechanism of control the steering wheel, the brakes, and the accelerator. Rather they are advice that overrides some of the algorithms. For instance, "steer out into the next lane and go around this truck" as the human realizes that the truck is just not going to move (see an anecdote below on the first night I took the new Waymo taxis in San Francisco (I had previously last ridden a Waymo in 2012 in Mountain View)). Why is this difference important? One of the motivations for self driving cars was that the economics of taxis, cars that people hire at any time for a short ride of a few miles from where they are to somewhere else of their choosing, would be radically different as there would be no driver. Systems which do require remote operations assistance to get full reliability cut into that economic advantage and have a higher burden on their ROI calculations to make a business case for their adoption and therefore their time horizon to scaling across geographies. But wait, you might say, isn't that electric car company that used to be based in California and is now based in Texas going to roll this out imminently and have a fully digital taxi service. They demoed it on a Hollywood movie studio lot just this year, and the cars were painted gold. Hmm. The location of the demo and the fact that the cars, even down to the tires, were painted gold tells you everything you need to know. Both the cars and the humanoid robots at that event were presented as autonomous but in reality they were all tele-operated directly by people (see below in the humanoid section for more details). And that same electric car company is actively hiring people into paying jobs as remote operators. There was a reasonably balanced appraisal from Reuters just after the event, though it does not go into details of the demos. Here is a direct quote from the story: "We do expect to start fully autonomous unsupervised FSD in Texas and California next year." Musk said. The astute reader will note that this is the 11^th year in a row that the CEO of Tesla has made this prediction of the same milestone happening the next year. We can admire the consistency. Actual self-driving is now generally accepted to be much harder than every one believed. The reason that this bait and switch is important to understand is that the promise of inevitable fully self driving technology upended a historical way that new transportation systems have been adopted. In the past whenever we have introduced new transportation mechanisms there have been large investments in infrastructure and that infrastructure is shared and used by everyone. The Romans built roads so soldiers and traded goods could travel long distances-in Europe those road networks are still the basis of today's road networks. When steam engine driven trains were the new transportation technology vast networks of rails were built allowing goods to move long distances in mere hours or days. When Ford started mass production of automobiles he built roads and the local governments followed and the the Federal government followed, and those roads are what we use today. Actual fully self driving cars promised that no infrastructure changes would be needed to revolutionize how vehicles would be controlled. Each individual vehicle would do what was needed all by itself. As sensors and networks got better there was no need for expensive new infrastructure because of this promise. The promise was false. If government and private partnerships in building smart roads, which was a hot topic in the 1990s. had continued, every one of us would now have smarter safer cars, but still with onboard human drivers taking over in many situations. But we would have had smart freeways where once you were on it your car would be self driving. The road would have had lots of sensors effectively shared across all cars, as that data would have been transmitted to all passing cars. It would have been a fraction of the cost per car compared to the sensing on today's almost but not really self driving cars like those of Waymo. And we would have had much more accurate congestion data where the root causes of local congestion would have been sensed with semantic understanding rather than just inferring it from the aggregate collection of location data from phones, individual cars, and historical data from roadside sensors. Instead we now have individual corporate actors using a mixture of partial self driving and remote human supervision. The big question is whether the economics of this works at scale, and whether the fake promises will drive out the human drivers in cheaper services and we'll all end up paying more. Will the level of hype we saw push our decentralized transportation system into the hands of a few wealthy companies, and in effect make it a centralized system where everybody has to pay private companies to be part of it? As a reminder of how strong the hype was and the certainty of promises that it was just around the corner here is a snapshot of a whole bunch of predictions by major executives from 2017. [Screenshot-2024-12-21-at-4] I have shown this many times before but there is one new annotation here for 2024. The years in parentheses are when the predictions were made. The years in blue are the years are the predicted years of achievement. When a blue year is shaded pink it means that it did not come to pass by then. The predictions with orange arrows are those that I had noticed had later been retracted. The prediction that Jaguar and Land-Rover made that they would have fully autonomous cars by 2024 did not come to pass, so I have shaded it pink, Note that every single blue year up until now is shaded pink, and that every one that is shaded pink has still not come to pass. None of the predictions that were out there in 2017 for the next few years have happened. None. There are three more for 2025, and I am sure that a year from now they will all be shaded pink also. One of the big selling points of self driving cars was that they would be safer than cars driven by humans. So far that is not holding up with real data. One electric car maker with self driving software had it disengage when it sensed there would be an accident, supposedly so that the human could take over in a split second. And then the company did not report the incident as the fault of the software as it was no longer controlling the car when the impact occurred. It was reported, and I had this experience myself in my last ride in a Cruise in 2023, that Cruise vehicles would freeze when an accident looked likely, and then not report it as their software's fault as the car was stationary and was hit by another car. In many reported cases, and in my case, simply continuing to move forward would avert any likely accident (fortunately for me the human driver of the other car slammed on the brakes and did not hit my robot vehicle). In this story from the Washington Post about Federal investigations into the safety incidents with self driving cars, they report that the companies involved claim they have vast amounts of driving on our roads under their belt. Not so. An industry association says autonomous vehicles have logged a total of 70 million miles, a figure that it compares to 293 trips to the moon and back. But it's a tiny fraction of the almost 9 billion miles that Americans drive every day. The relatively small number of miles the vehicles have driven makes it difficult to draw broad conclusions about their safety. To put that into perspective, the total number of miles driven by all autonomous (sort of) vehicles over the last decade is less than 1% of the miles driven by humans every day in the United States. It is a tiny, tiny portion. Take a look at this embedded video from the Wall Street Journal about investigations of crashes (many of which have been fatal) involving autonomous driving systems. From the audio: "The kinds of things that tend to go wrong with these systems are things like it was not trained on, pictures of an overturned double trailer. It just didn't know what it was. There were some lights there, but the lights were in unusual positions. A person would have clearly said something big is in the middle of the road. But the way machine learning works is it trains it on a bunch of examples and if it encounters something it doesn't have a bunch of examples for it may have no idea what's going on." [[My own take is that the fetish of end to end learning leads people to leave out well known algorithms that might solve many of these problems (e.g,, the incredibly simple time to collision algorithms based on looming). Yes, end to end learning made speech understanding systems better, but that does not mean it is the appropriate fetish to apply everywhere.]] Pro tip: Think about this history of industry prognostications about fully autonomous driving being just around the corner when you read today's prognostications about LLMs taking jobs, en masse, in the next couple of years, or humanoid robots being dirt cheap and being able to learn how to do any human manual task real real soon now. You know you have seen this movie before... My own experiences with Waymo in 2024 I have two sorts of experiences with Waymo vehicles. First, as a driver of my own vehicle and sharing road space with them every single time that I drive. And second, as a user of their ride service. The streets of San Francisco had been thick with Waymo vehicles with no driver in them especially in the second half of 2024. As I drive across the city every morning to head down to my robotics/AI startup half way down the peninsula I see them everywhere until I get on to 101. I see them in front of me and behind me and in adjacent lanes as I drive on multilane one way streets. Sometimes I see four of them in a single block. Twice I've seen four of them in a line, in my block and could see four of them in a line in the block ahead of me. When I am at four way intersections with no traffic lights I see them participating in the social ritual of taking your turn to drive through the intersection in the order you stopped, except when a pedestrian is crossing in front of you. They do that pretty well. They do less well when they accidentally get into a line of parents' cars snaking around a corner for school drop off or pickup. Over the last few months I have noticed that in general they are getting more aggressive about stretching the rules, just like people do. Otherwise human drivers (including me) take advantage of their politeness. That aggression is not always welcomed. One morning I saw a workman with a group doing some digging on a road, and holding a sign with SLOW on one side and STOP on the other side have to jump in front of a Waymo to get it to do what he was trying to tell it to do with the sign. STOP. It wasn't stopping for no stinking sign! The only time I have seen a Waymo go into reverse, ever, was when I was illegally driving the wrong way down a single lane street and we were heading straight at each other. As a rider I feel they are not quite aggressive enough with human drivers some time, so a ride in a Waymo takes longer than with an Uber or Lyft. It is hit and miss where they drop me off. Sometimes they take a place to pull over half a block from my house, even when it is raining. There is no way to adjust what they happen to decide that day, even though I know that they will always be able to pull in right in front of my house. The first time I took a Waymo this year, on the way home it picked me up at a restaurant and then was about to make a right turn. But at that corner there was an 18 wheeler with its lights flashing and surrounded by green cones. It pulled right in behind that truck and waited a long time before it drove forward. I am guessing a remote operator intervened told it to go around because eventually it pulled around it in the lane just to the left. Based on seeing Waymos interact with orange cones I suspect it would have done better if the cones had been orange rather than green. This easily illustrates that the learning that this robot does, and indeed any robot does, is nothing like the learning that people do (see my rant about the seven deadly sins and mistaking performance for competence in the section below on advances in AI and ML). I mostly feel safe when I am a passenger in a Waymo. Sometimes I don't feel that my driver of an Uber that I am taking rides with Uber that are not as safe as I would prefer. Self Driving Taxi Services There have been three self driving taxi services in the US in various stages of play over the last handful of years, though it turns out, as pointed out above that all of them have remote operators. They are Waymo, Cruise, and Zoox. Waymo and Cruise are similar in that they use conventional cars adorned with lots of sensors. Zoox has purpose built vehicles that have no steering wheel or pedals for brake or accelerator. Waymo and Cruise went for deployments in large parts of two or more cities and have had ride services callable by apps, just as one can do with Uber or Lyft. Zoox is smaller scale, much more restricted in geography, and really not comparable. At this time last year Cruise was in trouble has it had suspended all of its San Francisco operations under pressure from regulators after some bad accidents that happened in a way that never would happen for human driven cars. Briefly, their cars were getting hit at night by emergency vehicles with lights flashing as the Cruise cars crossed intersections. Human drivers see the reflections of lights from such vehicles flashing even if they don't see the vehicles themselves. The Cruise vehicles were only reacting to flashing lights that they could perceive directly. But the accident that tipped the scales was when a pedestrian crossing in front of a human driven vehicle was hit and went flying in the air landing right in front of a Cruise. The Cruise hit the person (who now disappeared from sight) as a human driver would most likely have done. But then it proceeded to drive 20 feet with the human underneath the vehicle being dragged along as it went into a mode where it was supposed to get off the road. A human driver would not have reacted that way to having been in a collision, even if it was not their fault. The hammer finally fell in December of 2024. General Motors shut down Cruise. The leading paragraphs from this linked story from the Wall Street Journal are: General Motors has scrapped its Cruise robotaxi program after nearly a decade and $10 billion in development, citing the time and costs needed to scale the business and rising competition. GM on Tuesday said it plans to realign its autonomous driving strategy and give priority to development of advanced driver assistance systems, which take over steering and other functions in certain situations and are common on new vehicles today. The automaker said it would continue to develop fully autonomous technology for personal vehicles, and build on the progress of its Super Cruise system, a hands-off, eyes-on driving feature that the company introduced several years ago. GM said it owns about 90% of Cruise and intends to buy out the remaining investors. It plans to combine the technical teams from Cruise and GM into a single effort to advance autonomous and assisted driving. "We want to leverage what already has been done as we go forward in this," Chief Executive Mary Barra told analysts on a call Tuesday. The Detroit automaker said it expects the restructuring to reduce spending by more than $1 billion annually after the proposed plan is completed, which is expected in the first half of next year. While there are 40 companies that have permits to test autonomous driving in California, alone, the demise of Cruise leaves just one company, Waymo, trying to make an actual go of a digital taxi service in the United States. They have an enormous significant lead over anyone else who wants get into this business and have spent billions of dollars (probably very much north of $10 billion) on this endeavor over the last 15 years. In an email they sent me a couple of weeks ago as a user of their services they reported that they provided 4 million customer rides in 2024. That is approximately 4 million more than any other company in the United States. Waymo Despite being so far out in front it has not been all smooth sailing for Waymo. Early in the year the operations center for Waymo somehow neglected to realize it was Chinese New Year in Chinatown in San Francisco. So Waymo vehicles were routed through that area on the biggest night of celebration. Any human driver would have realized that the streets, i.e., the street surfaces where cars usually drive, were completely packed with humans, no doubt some of whom were intoxicated as well as just being out having a good time. Not so the Waymo vehicles. They tried pushing through the very very dense crowds, no doubt annoying many people. And what do people have at Chinese New Year? Fireworks. So some revelers decided to push back on this robot car invading their space. Here are a couple of pictures of the results. [Waymo1] [Waymo2] Not pretty. And an example of how taking away people's agency is never a good idea for robots (see my second law of robotics). Throughout 2024 Waymo has been investigates for various accidents such as those described in this Wall Street Journal article. "Reports included collisions with stationary or semistationary objects, such as gates, chains or parked vehicles, according to the regulator." In the middle of the summer Waymo added a feature where they would honk their horns at cars in their way. But this backfired when hundreds of Waymos were coming back to their parking lot in the very early hours of the morning, and they started honking at each other and waking up human neighbors. Eventually that got fixed. In late September a motorcade for Kamala Harris in San Francisco was brought to a halt by a Waymo that stopped in the middle of California Street doing a U-turn in front of it. I'm sure this incident was of great concern to the Secret Service. Eventually a San Francisco police officer got into the car and drove it out of the way-this is shown in a video included with the story above. I do not know how the officer got access to the vehicle and whether Waymo remote operations were cooperating. More disturbingly humans outside the Waymos started harrassing humans inside them. The most concerning cases come from the realization that if a woman is in a Waymo at night she will be dropped off, outside, on a public road at the end of her journey with no option but to get out of the car where it has stopped. So groups of men have followed Waymos with women in them and then harassing the woman when she gets out. If she was driving her own car she might be heading to an off road parking space or she might choose not to stop if she knows she is being followed. There are no such options in a Waymo so taking a Waymo at night is less safe than other means of transportation-just follow it and eventually the preyed upon woman will have to get out. Here is a very recent disturbing story about this practice. Meanwhile Waymo managed to raise $5.6B to expand to new cities in 2025. It already operates in parts of San Francisco, Los Angeles, and Phoenix. The new money will let it expand to Austin and Atlanta in the United States and to start operating in parts of Tokyo in Japan. That is expensive expansion. Here is the question for the future of watered down remote monitored "autonomous" driving systems (let's call it "watered down autonomy"), and it is up to Waymo now. Can Waymo expand fast enough in these new markets in 2025 and take enough business from what is left of traditional taxi operators, along with those operating under the Uber and Lyft models, and do it in a way which is in sight of profitability, so that it has a case to raise the stupendous amounts of money needed to operate in all large cities in the US in the next 10 t0 20 years? If Waymo can not succeed at this in the next two years I think the idea of large scale use of watered down autonomy will be dead for at least a decade or two. Right now full autonomy everywhere is already dead. Prediction [Self Driving Date 2018 Comments Updates Cars] A flying car can There is a real be purchased by NET possibility that any US resident 2036 this will not happen if they have at all by 2050. enough money. Flying cars NET That would be about reach 0.01% of 2042 26,000 flying cars US total cars. given today's total. Flying cars reach 0.1% of US NIML total cars. This is a bit like current day HOV lanes. My bet is the left most lane on First dedicated 101 between SF and lane where only Silicon Valley cars in truly NET (currently largely driverless mode 2021 the domain of are allowed on a speeding Teslas in public freeway. any case). People will have to have their hands on the wheel until the car is in the dedicated lane. 20240101 This didn't happen in 2023 Such a dedicated so I can call it now. But lane where the there are no plans cars communicate anywhere for and drive with NET infrastructure to reduced spacing 2024 communicate with cars, at higher speed though some startups are than people are finally starting to look allowed to drive at this idea--it was investigated and prototyped by academia 20 years ago. 20240101 First driverless People may think this "taxi" service happened in San Francisco in a major US The pick up and drop in 2023, but it didn't. city, with off points will not Cruise has now admitted dedicated pick NET be parking spots, that there were humans in up and drop off 2021 but like bus stops the loop intervening a few points, and they will be marked percent of the time. THIS restrictions on and restricted for IS NOT DRIVERLESS. Without weather and time that purpose only. a clear statement from of day. Waymo to the contrary, one must assume the same for them. Smoke and mirrors. 20250101 Imminent dual use of personal cars was the carrot that got lots of people to pay cash when buying a Tesla for the Such "taxi" A key predictor here software subscription that services where is when the sensors would allow thei car to the cars are get cheap enough operate in this way. also used with NET that using the car Shockingly the CEO of drivers at other 2025 with a driver and Tesla announced in smoke times and with not using those and mirrors roll out of extended sensors still makes Cyber Cab in 2024, that geography, in 10 economic sense. the service would use major US cities specially built vehicles to be produced at some indeterminate late date. I got suckered by his hype. This is unlikely to happen in the first half of this century. It will be a very 20250101 slow start and roll Even the watered down Such "taxi" out. The designated version of this with service as above pick up and drop off remote operators is not in 50 of the 100 NET points may be used gong to happen in 50 biggest US 2028 by multiple vendors, cities by 2028. Waymo has cities. with communication it in 3 cities and is between them in currently planning on 2 order to schedule more in the US in 2025. cars in and out. Dedicated The geographies will driverless have to be where the package delivery NET roads are wide vehicles in very 2023 enough for other restricted drivers to get geographies of a around stopped major US city. vehicles. A (profitable) parking garage The economic where certain incentive is much brands of cars higher parking can be left and NET density, and it will picked up at the 2023 require entrance and communication they will go between the cars and park themselves the garage in a human free infrastructure. environment. 20240101 Looked like it was getting close until the dirty laundry came out. A driverless 20250101 "taxi" service Waymo now has a service in a major US This is what Uber, that looks and feels like city with NET Lyft, and this in San Francisco, 8 arbitrary pick 2032 conventional taxi years earlier than I and drop off NET services can do predicted. But it is not locations, even 2032 today. what every one was in a restricted expecting. There are geographical humans in the loop. And area. for those of us who use it regularly we know it is not as general case on drop off and pick up as it is with human drivers. Driverless taxi services operating on all Unless parking and streets in NET human drivers are Cambridgeport, 2035 banned from those MA, and areas before then. Greenwich Village, NY. A major city bans parking and cars with This will be the drivers from a NET starting point for a non-trivial 2027 turning of the tide portion of a BY towards driverless city so that 2031 cars. driverless cars have free reign in that area. The majority of US cities have the majority of NET their downtown 2045 under such rules. 20240101 This one looked pessimistic last year, but now looks at risk. There was a considerable slow down in the second Electric cars derivative of adoption hit 30% of US NET this year in the US. car sales. 2027 20250101 Q3 2024 had the rate 8.9% so there is no way it can reach 30% in 2027. I was way too optimistic at a time when EV enthusiasts thought I was horribly pessimistic. Electric car sales in the US NET make up 2038 essentially 100% of the sales. Individually owned cars can There might be some go underground small demonstration onto a pallet projects, but they and be whisked NIML will be just that, underground to not real, viable another location mass market in a city at services. more than 100mph. First time that Recall that a a car equipped variation of this with some was a key plot version of a aspect in the movie solution for the "I, Robot", where a trolley problem NIML robot had rescued is involved in the Will Smith an accident character after a where it is car accident at the practically expense of letting a invoked. young girl die. Electric Cars Last year US manufacturers pulled back on their planned production of EVs. In data from this report we can see that sales dropped at the start of 2024 but have now picked up again. 2022 2022 2022 2022 2023 2023 2023 2023 2024 2024 2024 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 5.3% 5.6% 6.1% 6.5% 7.3% 7.2% 7.9% 8.1% 7.3% 8.0% 8.9% There is steady growth in sales but my prediction of 30% of US car sales being electric by 2027 now seems wildly optimistic. We need two doublings to get there in three years and the doubling rate seems more like one doubling in four to five years. Note that some sources include hybrids and hydrogen powered cars in electric vehicles but I am using the battery electric vehicle (BEV) numbers. To see how the trends are across brands you can see a breakout for Q2 of 2024 here. There appear to be two main headwinds for BEV adoption. Firstly, if one doesn't have on property residential parking it is hard work in the US to find a place to recharge, and it takes hours for the charging to finish. This will stop many city dwellers from adopting. Secondly the increased tire wear adds up to real money. The maintenance requirements for BEVs are much less than for cars with an internal combustion engine. On the other hand tires do not last as long (I have had to buy four new tires in less than two years owning my first BEV), apparently due to the increased weight of the car. Flying Cars Flying cars are another category where the definitions have changed. Back when I made my predictions it meant a vehicle that could both drive on roads and fly through the air. Now it has come to mean an electric multi-rotor helicopter than can operate like a taxi between various fixed landing locations. Often touted are versions that have no human pilot. These are known as eVTOLs, for "electric vertical take off & landing". Large valuations have been given to start ups who make nice videos of their electric air taxis flying about. But on inspection one sees that they don't have people in them. Often, you might notice, even those flights are completely over water rather than land. I wrote about the lack of videos of viable prototypes back in November 2022. Nevertheless there have been wild predictions. I ended a longer version of this component in last year's annual review with: Also note the size of this vehicle. There are many fossil fuel powered helicopters that are much smaller. This is not going to be a personally owned vehicle for the masses. Don't hold your breath. They are not here. They are not coming soon. Nothing has changed. Billions of dollars have been spent on this fantasy of personal flying cars. It is just that, a fantasy, largely fueled by spending by billionaires. Robotics, AI, and Machine Learning So what happened in Robotics, AI, and Machine Learning this year? Many, many, many people got just a little bit over excited. That's what happened. There have been a lot of party tricks and it is the researchers who often play the tricks on themselves without realizing it. This is not new, none of it is new. But there are orders of magnitude more people watching it now, and more people are out to make a buck by being hypesters, promising riches to those who will invest in their irrationally overpriced companies. How could this be? We are seeing mass sinning, lots and lots of people committing some of the seven deadly sins of predicting the future of AI which I wrote about back in 2017 here (or here you can see a professionally edited version of that blog post of mine). Four of those seven sins seem most relevant to today's hyped up atmosphere around robotics, AI, and machine learning. [PerformCompet-1] [Magic-1] [Exponentialism-1] [Deployment-1] Here now are short descriptions of these particular four sins, edited down from my earlier much more detailed descriptions. Then I will weave them together to explain how it is still pretty much business as usual, and I mean that in a good way, with steady progress on both the science and engineering of AI. Performance versus Competence One of the social skills that we all develop is an ability to estimate the capabilities of individual people with whom we interact. We use cues from how a person performs any particular task to estimate how well they might perform some different task. We are able to generalize from observing performance at one task to a guess at competence over a much bigger set of tasks. These estimators that we have all inherited or learned do not generalize well to other creatures or machines. We are not good at guessing which smart things other species might be able to do, and we are not good at guessing what an AI system can do when we have seen it do a few tasks in a limited domain. We get it wrong all the time. Indistinguishable from Magic When people cannot explain how something works they cannot know its limits as they do not have any sort of model (nor have they seen enough examples of it before). Arthur C. Clarke said that any sufficiently advanced technology is indistinguishable from magic. In our minds UFOs can do all sorts of amazing things as we have no way of knowing their limits-they may as well be magic, And that is what they become in speculation about them. Isaac Newton spent half his working life on alchemy as he did not know that the nucleus of atoms were not subject to mere chemistry. He would have been just as ignorant of the limitations of an iPhone screen (different sort of apple...), despite his own ground breaking work in optics. Remember, he was a really really smart dude. But even he was not able to develop all the theories needed to understand the world around him, despite his successes with calculus and gravity and the makeup of white light. He attributed properties to chemistry that were way beyond its limits. Exponentialism We have just lived through sixty years of the most phenomenal growth of a technology in the history of humankind. It is the story of silicon-based computation. Everyone has some idea about Moore's Law, at least as much to sort of know that computers get better and better on a clockwork like schedule. This reality has trained people to think that probably a lot of other things in tech will change exponentially, especially when that thing has a strong computational component. The sin of exponentialism is to argue that some other process is going to follow a Moore's-like law when it is unwarranted to so argue. Moore's law worked for so long because in the starting technology of the 1960s the currents used to represent digital information were many many orders of magnitude beyond the minimal physical limit needed to determine whether they were present or not, and hence distinguish a 1 from a 0. Those currents could be halved many times without breaking physics limits. Speed of Deployment New technologies get deployed much more slowly than people imagine. Even software technologies. The old internet protocol, IPv4, can only address two billion, or 2x10^9, devices, which is way less than the number of people on our planet. A new protocol, IPv6, which can address more than 3x10^ 38 devices was meant to replace it over a two year period of dual use by about 2003. But in 2024 IPv4 was still there and carrying over half the world's internet traffic despite its inadequacies. Must functioning businesses that operate in the physical world are very averse to taking up new technology as it dramatically increases existential risk to their business. They must foresee immediate and incredibly high return on investment (ROI) to be tempted to move to new technologies. Even the military is slow to adopt new technologies. The US Air Force still flies the B-52H variant of the B-52 bomber. This version was introduced in 1961, making it 63 years old. The last one was built in 1963, a mere 61 years ago. Currently these planes are expected to keep flying until at least 2040, and perhaps longer-there is talk of extending their life out to 100 years. What does this all mean? Right now there is incredible hype for both Large Language Models (LLMs), and all their variations, and for humanoid robots, especially humanoid robots that are going to learn how to do things. The hype is driven by the four sins above. LLMs LLMs have proved amazing facile with language. They have been trained on pretty much all the text that is available on the Web and all the digitized historical books that exist. Miraculously LLMs seem to be able to infer a representation of some sort, that is somewhat independent of the particular human language that they read. So they are able to translate between human languages, and when you ask them just about anything they produce text in the language that you asked in, and that text often seems entirely reasonable and informative. I used the word "miraculously" as we do not really understand why they are able to do what they do. We, of course, know that the architecture for them is built around noticing correlations in vast amounts of text that connect some tens of thousands of tokens which are the components of words in each language that is digested. It is a surprise that they work as well as they they do, and produce coherent sounding language on just about any topic. Here is the original architectural diagram from the 2017 Attention Is All You Need paper: [attention] Each column from bottom to top is a pure feed forward network, with no search, no iteration, no conventional algorithm at all. There are inputs at the bottom and then layer upon layer of linear neurons that have numbers or weights stored in them that multiply and add their inputs and threshold that sum to provide an output. The detail in the architectural diagram is how the connections between layers are organized. On the left is an input or question, in a linear string of words, from a user. That gets injected half way up the network on the right and remains constant while a single iteration process runs. The stack on the right outputs a word (or token) and that gets fed back to the bottom of that stack, and a new token pops out the top. All the output tokens that have so far been produced remain in the right bottom input buffer as ordered input. What the network has been trained to do, is given the user input on the left, and what the network has output so far, choose a very likely next word, given the billions of examples it has seen in training. Some randomness is used to choose among a small number of very likely next words at each stage. There are hundreds of billions of weights that get learned and stored in the layers of network to act as multipliers for each individual input to each layer. So now us humans are faced with looking at this system running and our human nature just makes us commit the first two sins from above. It is in our nature and we cannot help ourselves. First, we see really impressive examples of responses to input questions, and if a human was giving those answers we would estimate that person to be quite clever and able to reason. Often though, because they have so many billions of examples on which they were trained LLMs are essentially looking up the question in the weights. The weight if gained from all of human knowledge that is out there on the network in language form. Invisibly the network is perhaps (but not in any intentional way) merging some similar questions, and then merging the answers which were already in the vast data that it has seen. But us dumb humans just think the damn thing is really really smart. Then, since we don't have a real explanation in our heads for what it is doing we start thinking it is magic, and that there is no real limit to what it is extracting from all that data (that it used a significant portion of the energy budget for many different countries to compute) and how general its capabilities will be. It becomes magic. And then researchers try to show that it can reason, that it has inferred a spatial understanding of the world, that language can be used to do all sorts of things that Moravec's paradox tells us it can't. There is a lot of magical thinking that humans do about LLMs. Of course it can diagnose diseases like a doctor talking about them. Of course it can teach a student as well as a human teacher. Of course it can program as well as a human computer programmer. It is magic after all. But in reality the fact that it is just picking likely next words means that in fact we can't trust its output. Some outputs are great. Some are pure confabulations (most people use the word "hallucinations" for this, but I prefer "confabulations"). And we do not know which we will get ahead of time, or more perniciously how much of each we will get, trustworthy pieces of output and confabulated pieces of output all jumbled together. Not to worry say the proponents, More learning will fix it. Fire up a nuclear power plant (I am not making this up-the tech companies are getting more nuclear power built or activated so that their LLMs can learn what a human learns using just 20 watts powering their brain; I am not confabulating this!!), and we'll feed it more data and it will become more trustworthy. It is magic after all. But the magic is not going as well as the proponents imagined and promised as this Wall Street Journal story explains. Their imaginations were definitely encourage by exponentialism, but in fact all they knew was that when the went from smallish to largish networks following the architectural diagram above, the performance got much better. So the inherent reasoning was that if more made things better then more more would make things more better. Alas for them it appears that this is probably not the case. But rabid exponentialists have not yet given up. Expect a bunch of VCs to adversely affect the growth of pension funds around the world as pension funds are a prime source of capital that VCs spend. More serious academics are working on boxing in the LLMs with more external mechanism beyond just feeding the output tokens back in as a linear string of input. Many of these mechanisms look a lot like more conventional AI mechanisms, and we will see where these additions prove to be useful, how much of the wheel will be reinvented, and how long (months?, years?, decades?) to get there. And the answers to those last questions will tell us how much sinning has been done by companies in predicting fast deployments. Back in rant at the beginning of this post I gave the example of I.B.M. and Watson and their completely optimistic predictions of how any problems of applying Watson (which seemed extremely competent based on its performance on live TV) to the real world would be solvable. The areas that it was predicted to be applicable came from magical thinking. Surely no one today could be as dumb as that big company was back in 2011. Surely not. No, not us smart inhabitants of 2025. Its us. We are nowhere near as dumb as them!! Humanoid Robots The other thing that has gotten over hyped in 2024 is humanoids robots. The rationale for humanoid robots being a thing is a product of the four sins above and I think way less rooted in reality than the hype about LLMs. In fact I think it is pretty dumb. [[I suspect many people will reason that I cannot have a valid opinion about this precisely because I happen to have built more humanoid robots than anyone else on the planet. So read ahead with caution.]] My first law of robotics states: The visual appearance of a robot makes a promise about what it can do and how smart it is. It needs to deliver or slightly over deliver on that promise or it will not be accepted. The first sentence describes, I think, what is sucking people into believing that humanoid robots have a big future. It looks like a human, so its performance will be like a human, so it will be competent like a human. It's the performance/competence sin without even waiting for the performance part! The second sentence describes how the humanoid fever will break, and how the hundreds of millions of dollars put into many of these companies (billions of dollars overall) will disappear. The puppets will not perform at acceptable levels. It is easy to see this as you hear all the things investors and CEOs of humanoid robots say they will be able to do. They have hardly even got to the lab demonstration phase. My third law of robotics is: Technologies for robots need 10+ years of steady improvement beyond lab demos of the target tasks to mature to low cost and to have their limitations characterized well enough that they can deliver 99.9% of the time. Every 10 more years gets another 9 in reliability. For real work, robots need to operate with four, five, or six nines. We are a long way from that. The zeitgeist is that we will simply teach the robots to do stuff and then they will be able to do it. BUT, we do not know yet whether that is going to work. In order for it to work you have to both collect the right sort of data and then learn the right things from that data. It is not at all clear to me that we know the answers to make either of those things true. I think it will be an active place for lots of good research for many years to come. There is an excellent survey paper of current research state of the art called Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes. Unfortunately I think the title of the paper is going to confuse many people. "Real-World Successes" to someone like me, who these days deploys robots that people pay for and that provide real ROI, sounds like it is about systems that have been deployed. But on reading the paper it turns out that they mean that it is learning and demonstrations done in a lab setting on physical hardware rather than just in simulations and simulators. And, to me the lab demonstrations are shakier (literally) than I imagined in my third law above. I think we are a long way off from being able to for-real deploy humanoid robots which have even minimal performance to be useable and even further off from ones that have enough ROI for people want to use them for anything beyond marketing the forward thinking outlook of the buyer. Despite this, many people have predicted that the cost of humanoid robots will drop exponentially as their numbers grow, and so they will get dirt cheap. I have seen people refer to the cost of integrated circuits having dropped so much over the last few decades as proof. Not so. They are committing the sin of exponentialism in an obviously dumb way. As I explained above the first integrated circuits were far from working at the limits of physics of representing information. But today's robots use mechanical components and motors that are not too far at all from physics based limits, about mass, force, and energy. You can't just halve the size of a motor and have a robot lift the same sized payload. Perhaps you can halve it once to get rid of inefficiencies in current designs. Perhaps. But you certainly can't do it twice. Physical robots are not ripe for exponential cost reduction by burning wastes in current designs. And it won't happen just because we start (perhaps) mass producing humanoid robots (oh, but the way, I already did this a decade ago-see my parting shot below). We know that from a century of mass producing automobiles. They did not get exponentially cheaper, except in the computing systems. Engines still have mass and still need the same amount of energy to accelerate good old fashioned mass. This Year's Prediction Update There is only one new comment in my robotics, AI and ML predictions table this year. There are a bunch of well funded new companies in the home robot space, and perhaps they will come up with new mobility solutions, which in my experience is the big blocker for home robots. Prediction Date 2018 Comments Updates [AI and ML] Academic Oh, this is rumblings about BY 2017 already the limits of happening... the Deep Learning pace will pick up. The technical press starts reporting about limits of Deep Learning, and BY 2018 limits of reinforcement learning of game play. The popular press starts having stories BY 2020 that the era of Deep Learning is over. VCs figure out I am being a that for an little cynical investment to here, and of pay off there NET 2021 course there will needs to be be no way to know something more when things change than "X + Deep exactly. Learning". 20240101 It definitely showed up in 2023. It was in Whatever this the public mind in turns out to be, December 2022, but was it will be not yet the big thing something that that it became during Emergence of the someone is already 2023. A year ago I generally agreed working on, and thought it would upon "next big NET 2023 there are already perhaps be thing" in AI BY 2027 published papers neuro-symbolic AI, but beyond deep about it. There clearly it is LLMs, learning. will be many and ChatGPT and its claims on this cousins. And, as I title earlier than predicted in 2018 it 2023, but none of was something already them will pan out. being worked on as the "attention is all you need" paper, the key set of ideas, was published in 2017. 20230101 The Turing Test was missing from all the breathless press The press, and coverage of ChatGPT researchers, and friends in 2022. generally mature Their performance, beyond the though not consistent, so-called I wish, I really pushes way past the "Turing Test" NET 2022 wish. old comparisons. and Asimov's 20240101 three laws as The Turing Test was valid measures largely missing from of progress in the press in 2024 AI and ML. also, and there was a story in Nature commenting on that. So yes, this has now happened. Despite some impressive lab demonstrations we Dexterous robot NET 2030 have not actually hands generally BY 2040 (I seen any available. hope!) improvement in widely deployed robotic hands or end effectors in the last 40 years. 20250101 A bunch of startups in the home robot space got significant A robot that can funding in 2024. Two navigate around Lab demo: of them are run by just about any NET 2026 ex-CEOs of large US home, with Expensive What is easy for companies: iRobot and its steps, its product: humans is still Cruise (and he was clutter, its NET 2030 very, very hard also an intern at narrow pathways Affordable for robots. iRobot after we were between product: already a public furniture, etc. NET 2035 company). So this one may be in play for a lab demo in the next few years if they have this as one of their goals.. A robot that can provide physical assistance to the elderly over There may be point multiple tasks solution robots (e.g., getting before that. But into and out of NET 2028 soon the houses of bed, washing, the elderly will using the be cluttered with toilet, etc.) too many robots. rather than just a point solution. A robot that can carry out the last 10 yards of Lab demo: delivery, NET 2025 getting from a Deployed vehicle into a systems: house and NET 2028 putting the package inside the front door. 20240101 One half of this happened this year. ChatGPT has been connected to A conversational Deployment microphones and agent that both platforms already speakers so you can carries long Lab demo: exist (e.g., now talk to it. and It term context, NET 2023 Google Home and does not fall into and does not Deployed Amazon Echo) so it recognizable patterns. easily fall into systems: will be a fast BUT the other half is recognizable and 2025 track from lab the half it does not repeated demo to wide have; it has no patterns. spread deployment. updatable memory apart from its token buffer of what it has just said. Long term context may be long term in coming. An AI system with an ongoing existence (no day is the I will need a repeat of whole new blog another day as NET 2030 post to explain it currently is this... for all AI systems) at the level of a mouse. This is so much A robot that harder than most seems as people imagine it intelligent, as NET 2048 to be--many think attentive, and we are already as faithful, as there; I say we a dog. are not at all there. A robot that has any real idea about its own existence, or the existence of NIML humans in the way that a six year old understands humans. A Parting Shot I recently read a research paper on humanoid robots working in built for human environments. It was based on the argument that the best form for a robot that is to operate in human environments is something tallish and skinny-ish, and probably dynamically balancing, with arms that can reach down to table tops etc., and with a sensor system that can look down from above, as that is what our human environments are optimized for. Here is the first paragraph of the paper: The past decade has seen an explosion of research in humanoid robotics. The stated motivations for this work have varied widely. Many teams have concentrated on bipedal locomotion, some have been interested in human level social interactions, understanding human intelligence, modeling human learning capabilities and others have been more interested in entertainment. Some humanoid robots have had manipulation capabilities on static humanoid platforms and some of that work is aimed at dexterity, plus there has been simple two armed grasping on mobile humanoid platforms. Overall there has been very little work combining dexterous manipulation with humanoid robots, static or mobile-much of that which has appeared, has been concerned with dynamic tasks like pole balancing and juggling rather than manipulation, or has used teleoperated manipulation. Apart from the weird references to pole balancing and juggling this all sounds pretty reasonable and consistent with what is happening today, and with recent history. In fact this is the very first paragraph of the very first paper in the very first issue of the very first volume of the International Journal of Humanoid Robotics. And it was published in 2004, with me as first author. Let me spell that out in case you thought there was a typo in the year. This is from a paper that I and my students and post-docs wrote in the year two thousand and four. Here is the beginning of the contents page for that first issue. [IJHR] You can download the text of that paper here. The journal is now in its 21^st year of operation, an on its 21^st volume of issues and papers. By the time this paper was written my research group at MIT had been working on and building humanoid robots for twelve years. This paper, about a robot named Cardea, was probably our sixth or seventh humanoid robot. [[In 2008 I started a company that built and shipped thousands of humanoid robots. The picture at the top of this post was taken in China with a line up of humanoids that we had built in Massachusetts and New Hampshire and sold to people in China (before a US initiated trade war with China put an end to it in 2018...irony can be personally hard to take at times...).]] The robot Cardea (Cardea was an ancient Roman goddess of door hinges and handles; these are still a challenge for modern robots...) was a two wheeled dynamically balancing robot that lived in a built-for-humans office environment. Cardea was able to open doors using existing door handles and then make its way through doors it had opened. Pro tip: Just because you heard about a new idea this last year or two doesn't mean that people haven't been working on that very same idea for decades. So temper your expectations that it must be about to transform the world. Ideas that transform the world take decades, or centuries of development, and plenty of people long before you have been just as excited about the idea and had thought it was on the verge of taking off. And none of us, including you and me, are likely to be special enough or lucky enough to come along at just the right time to see it all happen. Like all modern humanoid robots Cardea did not walk in a way that used passive dynamics to store energy, and basically modulate the behavior of a passive mechanism that had only low energy input, which is how all animals walk. So, like all modern mobile humanoid robots (and legged robots in general) when things were going awry its control algorithms tried to recover by pumping in large amounts of energy very quickly and sometimes that didn't quite work and the energy needed to go somewhere. Cardea could be a little dangerous in those circumstances, if it fell on you having just increased its kinetic energy. Even the spring based deployment system for its stick-like legs that were engaged when it realized it was going to fall could be dangerous. This is still a problem with all modern humanoid robots. That is why the tele-operated humanoids that were in the Tesla movie lot theater show a couple of months ago operated in two modes. When they all walked out the human guests were kept away from them. Once they stopped walking and were operating in a very different mode people were allowed to approach them, and then get fooled into thinking they were talking to an AI powered robot when they were really talking to a remote human operator. But the robot was no longer moving its feet, and no longer a source of physical danger as a result. Another pro tip: Don't stand anywhere near a walking or balancing wheeled humanoid when they are moving or doing any task. I have had some near misses for myself with my own humanoids twenty years ago and more recently with some of the humanoids from new start ups. And more generally never be below any sort of walking robot, no matter how many legs it has, when it is walking up stairs. HUMAN SpaceFLIGHT The numbers of flights in 2024 was not much different from those in 2023 (I neglected to include the flights by China last year). It does not feel like a golden age of human spaceflight, though there were other highlights from SpaceX. Orbital Crewed Flights Three countries put 28 people into orbit in 2024, the United States launched 16 people on five flights and Russia and China launched 6 people each with two launches. So there were nine crewed orbital flights total. Two were private and seven were government flights. The United States: There were four US flights to the International Space Station, starting with the private Axion-3 mission with a crew of four on January 18^th. The launch vehicle for this was a SpaceX Falcon 9, and the crew vehicle was a SpaceX Dragon. The remaining US flights to the ISS were paid for by NASA. Two of them were SpaceX flights, with four people on March 4^th, the Crew-8 mission, and two people on board Crew-9 on October 25^th. The remaining US flight to the ISS was the inaugural crewed flight of Boeing's Starliner, launched on June 5^th atop an Atlas V rocket with two people aboard. They are still stuck in space and will be for a few more months-see the section on Boeing below. The other US mission was also a SpaceX launch and vehicle flight, this time known as Polaris Dawn. It was the second mission paid for by billionaire Jared Isaacman, with him as commander. There was a former US Air Force fighter pilot as mission pilot and two SpaceX employees as mission specialists, giving a total crew size of four. They stayed aloft for five days, launching on September 10^th, This mission flew higher above Earth than any mission since Apollo 17, the last lunar landing mission, in 1972. Two of the crew "spacewalked" with their feet inside the Dragon capsule but with their bodies outside. This was the first private spacewalk ever. Now Isaacman has been tapped by the incoming US President to be the administrator of NASA. Russia: There were two Soyuz launches, each with three people, up and down, but different people coming back. The launch dates were March 23^rd and September 11^the. The six people that launched on Soyuz in 2024 were 3 Russian Cosmonauts 2 NASA Astronauts and one Belarusian commercial airline flight attendant who won a national competition with 3,000 applications. She was the only one not set for a long duration mission and was off the ground for slightly less than 14 days. So there were no space tourists per so, but the Belarusian flyer was most likely included as part of Russia's efforts to keep in good favor with Belarus which has aided it in its war in Ukraine, and was certainly not part of the regular scientific program of the ISS. China: There were two flights of Shenzhou (a larger more modern version of Soyuz) that were crewed in 2024. Both flights were to the Tiangong Space Station and both took along three Taikonauts, first on April 25^th and then on October 9^th. Both crews were assigned long duration missions and now the crews are overlapping previous crews at Tiangong so it is now being continuously occupied. The first handover this year took about five days and the second about three and a half weeks. Both times there were six Taikonauts onboard Tiangong at the same time. Suborbital Crewed Flights There have been two companies providing space tourism flights on suborbital flights. Blue Origin launches a capsule on top of a reusable rocket, New Shepard, and the capsule lands using a parachute and a brief rocket blast right before hitting the ground (similar to how Soyuz lands). Virgin Galactic has a winged craft which is carried aloft by a bigger a jet engined airplane, it separates at high altitude within the atmosphere and rockets into space. It flies back and lands on a runway. Both companies are run by billionaires who made their money in other businesses. Both billionaires have flown to space on their own craft. Both companies have aimed to have regular launches with lots of tourists, but neither has gotten to that scale and so far only a very small number of the many people who have paid a substantial deposit have been able to fly. Blue Origin had a failure with an uncrewed version of the vehicle in 2022 and only flew one flight in 2023 which was also uncrewed. This year they flew three crewed flights on May 19^th, August 29^th, and November 22^nd, each with six passengers (the system is automated and requires no pilots). In 2021 and 2022 they also had three flights, so there has now been nine crewed flights total. The first two took four passengers and the remaining seven have had six passengers, so altogether they have flown 50 people above the Karman line, 100 kilometers above Earth. This is not yet a regular cadence, nor a large scale tourist business. In 2024 Virgin Galactic had two flights, each with two crew from the company and four passengers. These flights were on January 26^th and June 8^th. Virgin Galactic flights are now on hiatus, awaiting a new bigger and better vehicle in about two years. Virgin Galactic has had a total of twelve flights since December 13th in 2018. Three have had two people on board and nine have had six people on board, for a total of sixty filled seats that have crossed the Karman line. The total number of different people is smaller as the two pilot seats on each flight have been occupied by a small number of people who have flown multiple times. So, in 2024 thirty people went on suborbital flights, and altogether there have been 110 people on these commercial suborbital flights. Space tourism on suborbital flights has yet to take off in a regular or scaled way. Prediction Date 2018 Comments Updates [Space] Next launch of people (test pilots /engineers) on a BY 2018 sub-orbital flight by a private company. A few handfuls of customers, NET 2020 paying for those flights. 20240101 There were four flights in 2021, three in 2022, and seven, five with customers on board, in 2023--all of them by Virgin Glactic. A regular Blue Origin did not fly in sub weekly 2023. At this point 2026 cadence of NET 2022 is looking doubtful for such BY 2026 regular flights every flights. week. 20250101 Now 2026 is looking impossible given the data from 2023 and 2024, and one of the two companies being on hiatus for all of 2025, and well into 2026. 20240101 Russia offered There were three paid paid flights to flights in 2021, and one Regular the ISS, but there each in 2022, and 2023, paying were only 8 such with the latter being the customer NET 2027 flights (7 Axiom 2 mission using orbital different SpaceX hardware. So not flights. tourists). They regular yet, and certainly are now suspended not common. indefinitely. 20250101 There were two paid flights in 2024. 20240101 Both SpaceX and Boeing were scheduled to have crewed flights in 2018. SpaceX pulled it off in 2020, Boeing's Starliner did not fly at all in 2023, but is scheduled to launch with people onboard for the first time in NET 2019 April 2024. Next launch BY 2021 20250101 of people BY 2022 (2 Current schedule The second company did into orbit different says 2018. finally launch humans into on a US companies) orbit in June 2024, so it booster. has happened three years later than I predicted and six years later than what had been promised when my prediction was made. Of course, everyone implicitly assumed that along with getting humans into space the companies would also be able to bring them back. Not so for Boeing. 20240101 Starship launched twice in 2023 but didn't get to orbit either time. This is Two paying going to be well over six customers go The most recent years later than the on a loop prediction has original prediction by the around the NET 2020 been 4th quarter CEO of SpaceX. Moon, launch 2018. That is not 20250101 on Falcon going to happen. The billionaire who signed Heavy. up for this and paid a hefty deposit in 2017 gave up waiting and cancelled the contract in 2024. This fantasy is over, for now at least. 20240101 I was way too optimistic, and bought into the overoptimistic hype of the SpaceX has said by CEO of SpaceX even though 2022. I think 2026 I added four years, Land cargo is optimistic but doubling his estimated on Mars for it might be pushed time frame. humans to NET 2026 to happen as a 20250101 use at a statement that it I can now call this as later date can be done, orbital mechanics and rather than for an Hohmann transfer windows pressing practical dictate that the cargo reason. would need to have been launched a few months ago for it to get to Mars in 2025. It has not been launched. Humans on Mars make Sorry, it is just use of cargo NET 2032 going to take previously longer than every landed one expects. there. It will be magical First for the human race "permanent" NET 2036 if this happens by human colony then. It will on Mars. truly inspire us all. Point to This will not point happen without transport on some major new Earth in an NIML breakthrough of hour or so which we currently (using a BF have no inkling. rocket). Regular I can't help but 20240101 service of be reminded of Calling this one 26 years Hyperloop NIML when Chuck Yeager early. As of today no-one between two described the is still working on this cities. Mercury program as in an operating company. "Spam in a can". Boeing's Starliner First announced in 2010 Boeing's Starliner was originally scheduled to fly a human crew in 2018. It carried out its second uncrewed flight in May 2022, and finally did make its first crewed flight on June 5^th. The crew of two docked with the ISS, but there were problems with multiple gas thrusters for fine motion during the docking. The original plan was that the crew would stay on the ISS for about a week and then return to Earth for a touchdown on to hard soil (as all Russian and Chinese crewed missions end along with all Blue Origin sub-orbital flights). The option of that return was considered, but the thrusters were on a section of the vehicle which is discarded along the way before the landing so there was no possibility of getting a look at the hardware back on Earth. So a program of tests while docked to the ISS was started delaying the crew return. Eventually it was decided that it was too risky for the crew to return on the craft and so it returned empty on September 7^th, landing in New Mexico. As it happened, although there were more anomalies with the thrusters the crew would have landed safely had they been on board. Now the crew was stranded in space with no designated ride home. It was decided to remove two crew from the Crew-9 launch and have the Starliner astronauts, Barry Wilmore and Sunita Williams, fly back on that SpaceX Dragon with the other two, which after additional delays is now scheduled to happen some time in March 2025. Their one week visit to the ISS will have stretched out to nine months by then. Boeing has committed to fixing the problems with Starliner. The boosters that it uses are no longer being built, but there are five existing ones reserved for the five additional contracted flights that Boeing has with NASA. They are supposed to happen once per year. We do not know at this point, but I think it would not be a huge surprise if Starliner never flies again. SpaceX Falcon 9 Once again the Falcon 9 launch system has broken all sorts of records for number of launches and reuse. During 2024 there were 132 single booster launches. For two of those flights no attempt was made to recover the first stage (there is a performance penalty for the primary payload in order to recover the first stage). One attempted recovery failed when the booster (on its 23^rd flight) caught fire as it landed on the recovery barge. Another booster has since flown a total of 24 times. In terms of mission success all but one of these flights succeeded; one failed when the second stage failed during re-ignition for adjusting the orbit. There were also two Falcon Heavy, the three booster version, launches, both of which succeeded. One of the had successful landings for the two side boosters, but there was no attempt to recoer the central booster on that flight and no attempt to recover any of the three boosters on the other Heavy flight. This brings the total number of launches of the single booster version to 417 along with 11 launches of the three booster Heavy version. These numbers are way beyond the number of launches for any other orbital booster. Additionally it is the only flying orbital system that is reusable at the moment, though Blue Origin and Rocket Lab both plan on joining the club soon. It is worth, once again, looking at how long it has taken to get to a total (across both single booster and Heavy triple booster versions) of 428 launches, with only three failures to deliver the payload to where it was intended to go. The first launch occured in June 2010, and there were a total of 4 launches in the first three years. The first successful booster recover happened on the 20th flight, in December 2015, five and a half years in. The first reuse of a booster occured in 2017, in the 8 ^th year of the program. Since 2021 there has been a steady increase in the number of launches per year, Year # of launches 2010 2 2011 0 2012 2 2013 3 2014 6 2015 7 2016 8 2017 18 2018 21 2019 13 2020 26 2021 31 2022 61 2023 96 2024 134 SpaceX had previously gotten satellites to orbit with its first rocket, the Falcon 1. Falcon 9 has been a spectacular success. But it was not instantaneous. It took time to build from the cadence of launches, about 10 years before the hockey stick curve showed up. Deployment is never sudden but comes after a long build. SpaceX Starship Starship is SpaceX's superheavy two stage rocket, designed to put 150 tons of payload into orbit, but also be able to go to the Moon or Mars. There is the booster which is designed only to work in Earth atmosphere with 33 Raptor engines both to get the second stage high enough and fast enough and to let the first stage have a controlled return to the launch site. The second stage, called Starship, is both a booster and the payload. It has three Raptor engines and three Raptor vacuum engines. The Raptor engines are designed to get the Starship into orbit after the first stage drops away, and to guide the Starship as it returns to its Earth launch site. The Raptor vacuum engines are meant for breaking out of Earth orbit and going to the Moon or Mars, and to do soft landings on those two bodies where there is no or almost no atmosphere. In 2024 SpaceX made steady progress with four launches of the two stages coupled together. The first two launches lead to both stages blowing up. The third and fourth launches were a big improvement. As with earlier flights they launched from the coast of Texas. In both cases the second stage did a reentry burn on it first orbit and then did a soft landing in a target zone in the Indian Ocean. In the third flight the main booster returned to the launch site and hovered next to the launch tower betweeen two giant arms which then captured it and the engines shot down successfully. It was sifficiently damaged during flight however, that it was not reusable. In the fourth flight there were health anomalies to the first stage was ditched in the Gulf of Mexico. On the fourth flight there was both less heat shielding and much less damage from heat during reentry. This is definite forward progress. But it is still quite a long way from both being operational and both stages being reusable. And it is even further away from being human rated. This is the vehicle that the CEO of SpaceX recently said would be launched to Mars and attempt a soft landing there. He also said that if successful the humans would fly to Mars on it in 2030. These are enormously ambitious goals just from a maturity of technology standpoint. The real show stopper however may be human physiology as evidence accumulates that humans would not survive three years (the minimum duration of a Mars mission, due to orbital mechanics) in space with current shielding practices and current lack of gravity on board designs. Those two challenges may take decades, or even centuries to overcome (recall that Leonardo Da Vinci had designs for flying machines that took centuries to be developed...). The President of SpaceX may be taking a leaf out of the CEO's always overly optimistic predictions. In November she said "I would not be surprised if we fly 400 Starship launches in the next four years". Looking at the success of Falcon 9 it is certainly plausible that I may live to see 400 Starship launches in a four year period, but I am quite confident that it will not happen in the next four years (2025 through 2028). One more thing. Back when I first made the predictions there had been an announcement by the CEO of SpaceX that in 2018 the company was under contract to send a very rich paying customer in a trip around the moon in 2018, launched on a Falcon Heavy. I was completely skeptical. Over the years the date got pushed back and pushed back, and the proposed flight vehicle was changed to be Starship. As we all know the flight of the Japanese billionaire around the Moon still hasn't happened. In 2024 Yusaku Maezawa finally gave up waiting and cancelled the contract. NASA Artemis NASA's plan is that the second Artemis mission, using the Orion Capsule, Artemis II, will fly to the Moon with four people aboard, the first crewed Artemis flight. An uncrewed flight of Orion around the Moon flew in 2022. The crewed flight was scheduled to launch in May 2024, but it was first delayed by six months and then a little more and in the last year it has slipped another full year. It is now scheduled to fly in April 2026. Artemis III was scheduled to launch in 2025 with a return to the surface of the Moon. However that relied on using a Starship (itself refueled in LEO by 14 (yes, fourteen!!) other Starship launches) to land there. No one any longer believes that schedule, and willlikely delay a few years, given where Starship is in its development and current capability. The officieal schedule says mid 2027, but that seems unlikely. You can find the architecture of the Artemis III mission at this website. Blue Origin Orbital BE-4 Engines and New Glenn The suborbital tourist flights that Blue Origin operates are not its main business. It has ambitions to compete head to head with SpaceX. Another billionaire vs billionaire competition. It has developed the BE-4 engine designed to fly 100 times, and to power the first stage of its massive New Glenn rocket (see below). But in the meantime it has started selling the BE-4 to ULA (United Launch Alliance) to power their Vulcan Centaur heavy launch vehicle. It's first stage uses two BE-4 engines, along with a variable number of solid fuel strap ons. Vulcan Centaur flew two times in 2024 and the BE-4 engines worked perfectly both times, on January 8^th and again on October 4^th. This is a solid validation of the engine's capabilities. Blue Origin's own first orbital class rocket, New Glenn, is massive, and comparable to the Flacon Heavy (three boosters) rather than the Falcon 9 in capability. It has been in development for a long time, but saw its first visits to a launch pad, fully stacked in 2024. The first stage uses seven BE-4 engines, and is intended to land on a barge and be fully reusable. The second stage uses two BE-3U engines, a variant of the single engine used on their New Shepard sub-orbital space tourism vehicle. There is a project underway to make a fully reusable version of the second stage. Launch seems imminent. Here it is at the launch pad in November 2024. [image] On Friday December 27^th, 2024, it was fully fueled in both stages and went through a countdown and fired its seven BE-4 engines for 24 seconds. Now it will leave the pad to have its payload installed. The launch could be as early January 6^th. The very first launch will be an all up affair, attempting to get something to orbit and land the booster on its first flight. This is a very different development approach to that used by SpaceX. Let's Continue a Noble Tradition! The billionaire founders of both Virgin Galactic and Blue Origin had faith in the systems they had created. They both personally flew on the first operational flights of their sub-orbital launch systems. They went way beyond simply talking about how great their technology was, they believed in it, and flew in it. Let's hope this tradition continues. Let's hope the billionaire founder/CEO of SpaceX will be onboard the first crewed flight of Starship to Mars, and that it happens sooner than I expect. We can all cheer for that. Categories: Dated Predictions Comment on this Cancel reply Your email address will not be published. Required fields are marked * [ ] [ ] [ ] [ ] [ ] [ ] [ ] Comment [ ] Name * [ ] Email * [ ] Website [ ] [Post Comment] [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] Post navigation Previous Previous post: Tips For Building and Deploying Robots (c) 2025 Rodney Brooks. All Rights Reserved.