Factory VLA Execution Checks

Factory VLA pilots need per-episode failure labels tied to production quality checks. The Siemens packaging study gives a useful starting schema for transparent accessory-bag insertion: contents above the product, multiple-bag grasp, incomplete insertion, and poor or failed grasp. The largest reported problem was bag contents left above the product, accounting for 65% of failed unconstrained episodes.

A manufacturing automation team could add these labels to every rollout review, then collect recovery episodes against the largest bucket after each evaluation round. The same review can track constrained and unconstrained trials separately, because the Siemens team first simplified the task and then removed constraints over later rounds. A cheap pilot is one packaging cell, one VLA policy, 30-bin-emptying style evaluations, and a rule that the next data collection round targets the top two failure categories.

Robot teams adapting a pretrained VLA to a new manipulation task can test a small online RL loop that edits action chunks while keeping the base policy intact. EXPO-FT trains a lightweight edit policy over a pretrained π0.5 policy, lets a Q-function choose the base or edited action, and stores human corrections to individual timesteps inside action chunks.

The practical trial is narrow: run the zero-shot policy, add sparse binary rewards and rule-based success detectors, collect online rollouts with operator corrections, then run 30 human-judged evaluation trials. EXPO-FT reports 30/30 final successes on each of 8 real-world manipulation tasks after an average of 19.1 minutes of online robot data. That result is a reason to test an edit-policy loop on precision manipulation jobs where full task data collection would take days.

A VLA deployment checklist should include an image-perturbation test that measures action changes, not only clean-task success. The adversarial reliability paper cites OpenVLA-7B falling from above 95% LIBERO success to under 5% under a 16/255 PGD image attack, and it frames the issue as physical action safety because the model output moves the robot.

A practical audit can run PGD and Square attacks on a held-out set, compare clean and attacked actions across short horizons, and compute the paper’s encoder-specific ceiling or head-agnostic robustness ratio. The paper says these diagnostics can be computed from at most 200 samples and reports zero bound violations across 320 validation cells covering OpenVLA, LIBERO suites, attack types, horizons up to 10, and two action-head designs. This gives robot integrators a small pre-flight test for camera-facing policies before floor trials with people nearby.