Embodied policies improve by preserving action-relevant state

Current images often omit the information needed to finish a physical task. POT-VLA keeps role-indexed 3D object records across walking, contact, occlusion, and recovery, reaching 71/80 real-world successes versus 39/80 for its matched baseline. FM-VLA instead compresses wrist force history into eight memory tokens; it averages 83.3% across three contact-rich tasks, compared with 27.8% for a memoryless policy. Together, the studies show that useful memory should preserve physical events and entities, not merely add more past images.

Several papers improve control by exposing structure that generic representations discard. Patch Policy retains dense Vision Transformer patches and reports stronger real-robot results than OpenVLA-OFT with 51.55 million parameters rather than 7.61 billion. SAGE predicts reachable latent subgoals before proposing actions, lifting horizon-150 success from 12.7% to 64.7% on PushT. GeoWorldAD similarly conditions driving plans on present and predicted 3D geometry, reaching 91.0 PDMS on NAVSIM v1. The common mechanism is a smaller, action-oriented search space—not scale alone.

Explicit intermediate computation is not itself evidence of robust control or causal understanding. In a VLA comparison, latent iterative reasoning fell to 14.8% success under Gaussian vision noise, while text chain-of-thought retained 92.7%; an adaptive attack also reduced a reasoning-based detector’s AUC from 0.996 to 0.493. A separate video study finds that generators can produce plausible continuations despite weak explicit causal perception. These results narrow the claim: reasoning must be tested at the point where it affects actions or predicted outcomes.