State representations and checks for embodied control

Robot safety and deployment teams should check action completion against independently refreshed physical state rather than a policy’s reasoning trace. POT-VLA shows how the same role-indexed 3D object records can condition actions and then be refreshed after execution to test containment, support, alignment, or handover relations. This is a more concrete monitoring surface than text-plan consistency, which fell to chance under adaptive attack in the VLA robustness study.

The build change is to expose object relations—and contact events where force sensing is available—as a small execution contract after every action chunk. The monitor should receive new sensor measurements rather than hidden states or generated explanations, and it should trigger re-observation or recovery when evidence is uncertain. A useful first test would corrupt vision, force, and policy internals separately and compare this state-based check with plan–action and action-anomaly monitors at matched false-positive rates; the result would show whether physical grounding adds independent safety information rather than another view of the policy output.

Engineers fitting manipulation policies onto latency-constrained robots should allocate tokens asymmetrically: preserve dense patches for the current image, but compress history into object and contact records instead of retaining past frames. Patch Policy’s compression ablation indicates that reducing a current observation from 256 patches to 64 or fewer damages precise control, while FM-VLA shows that long force histories can be compressed into eight tokens with only 3.3 ms of added latency. POT-VLA adds a complementary form of compact history through persistent role-indexed 3D object slots.

A fixed-latency comparison should pit this design against uniform visual-token compression and sampled-frame memory on tasks that combine precise alignment, occlusion, and repeated contact. The decision-relevant measurement is not only task success but failure type: current-frame compression should increase localization errors, whereas missing physical history should increase wrong counts, premature completion, and failed recovery.

Teams selecting world models for robot or driving planners should test whether predicted futures change correctly under action and scene interventions, not whether rollouts merely look plausible. SAGE shows that planning performance depends heavily on which action sequences reach the frozen world model: removing its ranking and CEM refinement cuts horizon-150 PushT success from 64.7% to 16.0%. GeoWorldAD uses future geometry to guide trajectory refinement, but its reported excerpt does not isolate present geometry from predicted future geometry. Separately, Thinking in Video finds that generators can produce plausible continuations despite weak explicit causal perception.

The evaluation change is a matched counterfactual suite: hold the observation fixed, alter one candidate action or one causal scene variable, and score whether predicted object geometry, collisions, and candidate ranking change in the correct direction. Running this before a larger physical trial would distinguish a model that supplies action-relevant dynamics from one whose plausible futures contribute little beyond a strong proposal generator or current-scene geometry.