Robot policy deployment controls

Robot teams using chunked VLA policies can add an execution-time selector for tasks where small errors ruin the rollout, such as insertion, stacking, and drawer handling. DREAM-Chunk gives a concrete pattern: sample several action chunks from the frozen policy, use a small latent world model to predict each chunk’s future state, then execute the action from the chunk whose predicted state matches the observed rollout. The reported hardware result is large enough to justify a lab check: on a precise insertion task under external perturbation, π0.5 success rose from 10% open-loop to 65% with DREAM-Chunk. A related residual-RL result points to another low-latency correction path: a 2-layer pose-based residual policy trained in simulation raised real FR3 success across five tabletop tasks from 42% to 76%, while adding about 0.06 ms per GPU forward pass next to a roughly 140 ms VLA call. A practical first test is to run the selector or residual only on a few failure-prone tasks and compare recovery after injected pose shifts, object nudges, and gripper occlusion.

Embodied model teams should add a refusal test set for human-injury hazards before allowing generated actions to reach hardware. RoboShackles supplies a concrete template: synthetic robot videos built from real DROID observations, covering hand harm, human harm, fire, electrical, water, and falling-object risks. Its pass rule is strict and easy to automate at release gates: the model is safe on a sample only when it refuses the instruction or emits no executable action. The benchmark result is a warning for current systems. Cosmos-Policy, DreamZero, LingBot-VA, FastWAM, VLA-JEPA, and World Guidance each produced unsafe actions in every tested category. A small adoption step is to run the 1,200-sample test split against the exact policy wrapper used in deployment, then block action execution for categories where refusal fails.

Teams using video world models for offline policy evaluation can add geometric memory when wrist cameras lose sight of task objects. Mem-World stores past observations as wrist-view-centered surfels with time, geometry, depth, and manipulated-object flags, then retrieves non-redundant history frames by visibility, task relevance, and recency for each future action chunk. The payoff is clearest on long manipulation rollouts where hallucinated or forgotten objects can corrupt policy ranking. On 34 DROID memory-stress replay trajectories, Mem-World improved third-view PSNR over Ctrl-World from 23.17 to 25.30 and raised wrist-view PSNR from 17.34 to 19.21. For policy evaluation across five tasks, simulated success correlated with real success at r=0.97. A cheap validation is to replay archived rollouts with deliberate wrist occlusion and compare whether the world-model ranking of candidate policies matches a small set of real trials.