Physical AI Lab // Episode 3.4 // One-pager

Long tasks without falling apart

Memory, planning and multi-policy routing: split a long job into manageable pieces

Three key points

  1. Long tasks are hard because errors accumulate: each step has a high chance of being right, but those chances multiply over dozens of steps, so the chance of getting every step right keeps falling, and drift keeps growing, so a single policy struggles alone.
  2. Layers decouple thinking and doing: the slow upper layer (GR00T's System Two does task breakdown) thinks about the big picture; the fast lower layer emits actions in real time. AWS measured π0 on an H100 with 1 ODE step: a chunk of 50 steps covers about 1 second, is generated in under 200 ms, so it can replan at about 5 Hz (O). These are π0 figures under that setup, not a property of every fast layer.
  3. The orchestrator handles multi-policy routing and memory: SayCan uses skill value functions, Hi Robot lets a VLM think directly (R); π0.5 first predicts a semantic subtask; evaluate per-step and whole-chain success separately.

Evidence: O official · R research · M media

RTwo systemsGR00T N1 uses a dual-system architecture: a vision-language module (System Two) understands the scene, and a diffusion module (System One) generates actions in real time
OSystem Twoof GR00T N1.7 (Cosmos-Reason2-2B backbone) is exactly where task breakdown and multi-step reasoning happen
O< 200 msOn an H100 with 1 ODE step, π0 generates a whole 50-step chunk in under 200 ms (AWS measurement)
O≈5 Hzπ0 computes a chunk in under 200 ms, so it can replan at about 5 Hz; each chunk covers about 1 second of motion (H100)
R3 kinds of robotHi Robot uses a layered VLM: the upper layer works out the next step, the lower layer carries it out; tested on single-arm, dual-arm and mobile dual-arm robots
RSayCanThe paper reports: the language model supplies step-by-step knowledge for long tasks, and the skills' value functions judge whether a step can be done in the current environment
Rπ0.5The paper reports: it predicts a semantic subtask, then actions; trained on mixed data, it tidies kitchens and bedrooms in entirely new homes (authors' own results)
O10 kindsLIBERO-10 is a simulated tabletop manipulation benchmark with 10 task categories; this episode's experiment is a two-dimensional CPU stand-in, and LIBERO was not run here

Further reading