Catching Physical Hallucinations: A Causal Execution Layer for VLA and Autonomous Driving under Distribution Shift
Year: 2026
Type: preprint
Abstract: Vision-language-action (VLA) policies are trained to produce actions, not to know when their actions stopped being true. Their one self-assessment is a scalar confidence, and a confidence is a function of the model while physical drift is a function of the world; the two can disagree without either being wrong. We describe a causal execution layer (CEL)—in one phrase, a causal auditor of the policy’s own physical claims—that replaces the question “how sure am I?” with “did the world do what I said it would?”. It never filters, clips or blocks an action: it audits the claim attached to one. Before each action chunk the layer registers the physical consequence its world model predicts; after the chunk it measures the consequence, quantises it in the plan’s own frame, and offers it to a hash-linked gateway. A matching digest earns an execution warranty—evidence, bound to this attempt, that the commitment was redeemed. A refusal is written to the ledger and the plan is re-derived from the observed state rather than retried. On libero_spatial task 2 (200 episodes, two arms, identical policy) the layer raises success from 51.0% to 100.0%; under a 101–108 mm displacement of the target plate it is 100.0% against 0.0% (p < 10⁻²⁴); and the confidence baseline refuses 34 of 84 actions that would have succeeded. Total latency rises 1.7% while decision latency falls 19.3%, because the layer never pays for a blind retry. The layer is not specific to an arm. We move it to an autonomous-driving stack, where the drifted quantity is the friction of the road rather than the pose of an object and no camera observes it: on 100 episodes of a 280 m urban route, a confidence-gated planner ends 15 of 30 wet-road episodes in a pedestrian collision and the layer ends none (Fisher 5.8 × 10⁻⁶); in a worn-steering channel where both arms reach the goal every time, 96% of the baseline’s episodes arrive over the centre line against 0%. The same layer also recovers the physical parameters it was never given, µ̂ = 0.350 and τ̂a = 0.400 s, from its own saturated and stepped responses. We also report a defect in our own first implementation: the gate was a set-membership test over all previously registered digests, so a scene that regressed to an earlier state was accepted as evidence that the action had worked. Three episodes passed every witness check with the manipulated object 245–269 mm from its target. A hash binds content; causation is a relation between a prediction and a particular attempt. Binding a monotonic ordinal into every registration closes the replay, and we argue this is a general requirement for any physical witness, not a bug fix.
Author: Liao, Zhizhi(廖至智)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225326400