Robot-Side VLA Release Checks

A manipulation team that already has an imitation-trained VLA can add a bounded online fine-tuning stage on the target robot before release. The useful build is a rollout runner that keeps the demonstration buffer, collects on-policy attempts, warms up the critic on mixed offline and rollout data, and filters candidate actions by value before updating the policy.

FORCE gives a concrete test case for this workflow. On six real-world Franka tasks, its reported average success rose from 45.0% under behavior cloning to 98.3% after online fine-tuning, with average execution steps falling from 112.8 to 38.9. The reported online stage used no human intervention. A practical first check would run the procedure on one high-value station task and compare behavior cloning, critic warm-up without actor updates, and full value-filtered updates under the same rollout budget.

VLA deployments with moved cameras, calibration drift, or changed end effectors need a small setup-identification step in the runbook. The build is direct: before the task, execute a few safe target poses, record the start image, action, and resulting image, then pass those clips as cached context for the task policy. The same log format can store action-consequence triplets during execution, so the policy can keep seeing how its own commands change the scene in the current setup.

ICWM reports gains under unseen camera viewpoints using task-agnostic probing clips, with no test-time weight updates or task demonstrations. Reflective VLA reports that full observation-action-consequence triplets beat observation-only and action-only context under camera and robot-calibration shifts. G3VLA adds a complementary engineering route for multi-camera cells: inject intrinsics and extrinsics into visual tokens through ray embeddings, PRoPE, and cross-view fusion while keeping the action pathway unchanged.

VLA test suites should log whether each completed rollout stayed safe through the whole motion. The concrete tool is a safety ledger for every episode: safe success, unsafe success, safe failure, unsafe failure, cumulative safety cost, and risk exposure time. This matters for release decisions because a robot can finish the instruction after brushing a hazard, entering a restricted zone, or colliding with a nearby object.

ForesightSafety-VLA shows the shape of the measurement. Across completed baselines, unsafe success appears even for the strongest listed policy: OpenVLA-oft reports unsafe success rate 0.06 and cumulative safety cost 0.18, while weaker baselines show higher unsafe shares among successful episodes. LIBERO-Safety adds a practical source of test cases: 75 tasks across physical and semantic safety suites, plus 19,664 screened collision-free demonstrations generated with keyposes and CuRobo collision checks. A small deployment version can start with the hazards already present in the workcell and add safety predicates to the rollout logger before expanding the task set.