Physical AI Lab // Episode 0.3 // One-pager
See, think, act
One second of a robot's life, in slow motion
Three key points
- See: both the image of the outside (a public dataset records 15 frames a second, R) and "itself": the position and pose of its hand (proprioception).
- Think: two systems, a slow one that thinks it through and breaks the task down, and a fast one that produces actions (NVIDIA GR00T N1.7, O); in one paper, the robot (π0) thinks again after every 25 steps, about every half second (R).
- Act: the control interface exchanges commands at a thousand times a second (one company's FCI, 1 kHz, O), so it can fine-tune in time. Being both fast and accurate is still a long way off.
Evidence: O official · R research · M media
| R | 15 frames/s | recording rate of a public robot dataset (images and actions both logged 15 times a second) |
| O | Slow and fast | the dual-system design of GR00T N1.7, an open humanoid robot model: high-level reasoning kept apart from low-level action |
| R | 25 steps / half a second | π0, on robots running at 50 Hz, re-runs inference after executing every 25 actions, about every 0.5 s (paper) |
| O | 1,000 per second | a robot company's control interface (FCI) sends and receives low-level control commands at 1 kHz |
Further reading
- DROID: a large-scale in-the-wild robot manipulation dataset (arXiv 2403.12945)
https://arxiv.org/abs/2403.12945 - π0: a vision-language-action flow model for general robot control (arXiv 2410.24164)
https://arxiv.org/abs/2410.24164 - NVIDIA GR00T N1.7 (Hugging Face blog)
https://huggingface.co/blog/nvidia/gr00t-n1-7 - Franka Control Interface (1 kHz real-time control)
https://franka.de/tmr