CausalWM world model predicts motion and 3D geometry, then turns it into robot policies
arankomatsuzaki · x · 2026-10-14
Following up on the robot demo, Komatsu-zaki reveals the learned policies come from CausalWM, a world model that predicts motion and 3D geometry before generating future frames; an action module on the same backbone converts predictions into executable policies.
The architecture diagram shows a causal agent whose task state updates continuously from user requests, multi-camera observations, and verifier feedback (e.g. verify(grasp)→false, joint-limit errors). It loops through stage identification, tool selection, outcome verification, and retry/replan, with a unified tool interface mixing the world-model policy with rule-based functions, SLAM navigation, and verifiers — needing only a few demonstrations.
Related event: Robot tablecloth demo combines rules with world model(2 posts)→
More from Embodied
- Visor: open-source Android app replaces the Meta AI app on Ray-Ban Meta glasses with your own model — jedisct1 · 2026-10-14
- 304 robotics papers hit arXiv in a single day as the field explodes — DJiafei · 2026-10-14
- NTU's COAP: code replaces LLMs in the robot control loop, hitting 75% on 42 bimanual tasks — liuziwei7 · 2026-10-14
- WAVE-0 adds a 257M world model to VLA, lifting real-world success from 38.3% to 73.3% — Haomin Wang · 2026-10-14
- Robot tablecloth demo works via staged repair loop: rule-based checks plus learned policies — arankomatsuzaki · 2026-10-14
- 99%+ data gap: Sophia Capital partner on embodied-AI investing — 创业邦 · 2026-10-14