Dynamics grounding

Haoyang Weng's essay (Oct 2026) broke my assumption that the gap is mainly inference speed. Spatial grounding is not dynamics grounding. Running Astra at 50 Hz just gives you 50 uninformed guesses a second, because closed-loop control is a learned mapping from feedback to correction, and you only get that by acting on your own body and feeling the consequences. Reading every book on bikes doesn't teach you to ride. His stack: S1 (LLM/VLA, ~1–5 Hz) produces interaction intent, and S0 (the low-level controller) executes it. Today's S0 is usually a free-space motion tracker that is blind to the scene and never trained with object loads, so most of S1's intent is lost through a motion-only interface. What he wants instead is one general S0 trained with diverse object dynamics, taking scene input (point clouds/RGB/depth) and a richer interface: object motion as 3D flow, plus explicit contact/force targets, since the same motion can mean wiping vs hovering, or sitting vs squatting. Steerability helps S0 learn the distinctions and S1 express them. Feeding (s, a, s′) to Astra would let it explain what happened, but deriving the correction means running dynamics in its head, which is MPC by another name, and training it on lots of those transitions is just training a controller again.