Beyond motion tracking: why robots still need closed-loop S0 controllers

animesh_garg · x · 2026-11-22

Animesh Garg shares Haoyang Weng's research note arguing that LLMs/VLMs like Astra can serve as upper-level policies (S1) generating interaction intent, but cannot directly control robots. Spatial grounding isn't dynamics grounding: producing commands at 50 Hz doesn't teach a policy to adjust as contact, loads, and disturbances change. True dynamics grounding requires closed-loop interaction experience — acting, sensing effects, adapting — which supervised learning struggles to provide. Robots still need dynamics-grounded S0 low-level controllers with expressive interfaces that let S1 steer physical interaction precisely.

Original post →

More from Embodied

Embodied channel →