Three-layer robotics stack — VLA world model plus RL whole-body controller — is already SOTA
zhaoran_wang · x · 2026-11-02
@archiexzzz predicts future robot architecture in three tiers: System 2 uses LLM agents for long-horizon planning, task decomposition, and closed-loop replanning; System 1 is a hybrid VLA-WAM policy jointly modeling future world states and actions, outputting action chunks or motion tokens; System 0 is a small (10M-100M) learned whole-body controller trained with RL and human motion priors. @JasonMa2020 notes this is already the present for SOTA robotics. The attached figure shows a concrete instance: the Taku semi-humanoid (21-DoF upper body + mobile base) running an RL whole-body controller at 100 Hz, a Dyna-2 world-action model at 5 Hz, and a VLM workflow orchestrator on top.
More from Embodied
- CorebrumOS Physical AI OS now runs on NVIDIA Jetson with 8 agents — chrismatthieu · 2026-11-02
- Hacked Feetech servo firmware cuts SO-ARM robot arm error from 35mm to 0-2mm — micoolcho · 2026-11-02
- Homebrew 4DGS pipeline with a 10x Raspberry Pi global-shutter camera rig — willeastcott · 2026-11-02
- Apple's touch-screen MacBook Pro with M7 Pro/Max chips and new AI hardware coming next year — petefang · 2026-11-02
- Rocket Lab CEO Peter Beck Buys and Flies a Jetson ONE Personal eVTOL — BrettKrieger12 · 2026-11-02
- UW–Madison ECE opens faculty hiring in Physical AI at all ranks, deadline Dec 14 — prof_kamilov · 2026-11-02