[HN Gopher] Real-time action chunking with large models
___________________________________________________________________
Real-time action chunking with large models
Author : pr337h4m
Score : 65 points
Date : 2025-06-17 19:39 UTC (1 days ago)
(HTM) web link (www.pi.website)
(TXT) w3m dump (www.pi.website)
| jauntywundrkind wrote:
| Anyone have good intro recommendations for VLAs?
| lachyg wrote:
| (I work at Pi.)
|
| We open-sourced Pi0 (referenced in this post):
| https://github.com/Physical-Intelligence/openpi
| fennecbutt wrote:
| Alright, I'm building the robot project I was putting off. This
| is so fucking cool.
|
| Excellent work!
| UltraSane wrote:
| I love the implications of a robot that can plug in Ethernet
| cables.
| meepmorp wrote:
| "Soon, a robot will fix the cables in the server room for me!"
| b0a04gl wrote:
| rtc handling 300ms+ delay and still pulling off tasks like
| plugging ethernet is kinda nuts. what i'm not getting is but
| how's it keeping the control loop stable without retraining? some
| sort of latent plan caching?
| kvablack wrote:
| It uses an inpainting algorithm (adapted from image generation
| literature) to produce future actions that are consistent with
| the current trajectory. It's sort of like warm-starting from a
| cached plan, although the plan isn't latent, it's directly in
| action space. Hopefully that answers your question -- there are
| many more details in the blog post and paper :)
___________________________________________________________________
(page generated 2025-06-18 23:00 UTC)