Robot Policy Reliability

VLA teams working on bimanual or contact-heavy manipulation should add failure and recovery episodes to the training buffer with separate labels, then evaluate policies with injected grasp errors. RePO-VLA gives a clear template: slice recovery segments out of full episodes, reset observation history for those segments, keep useful prefixes from failed rollouts with reliability decay, and assign low value to terminal drift. The deployment path is also practical because the reported policy uses a fixed high value condition, without an online failure detector or a hand-coded retry rule.

The first cheap test is a small adversarial suite around the failure modes that already break the robot: premature close, grasp slip, grasp position offset, and grasp orientation mismatch. RePO-VLA reports average adversarial success rising from 20% to 75%, and its FRBench-Sim data includes 23,453 bimanual episodes across 46 tasks, including 6,392 verified failure-recovery episodes. That scale is larger than most labs will copy at first, but the workflow is clear enough to pilot on a few high-value tasks before expanding collection.

Robot deployment teams can add an inference-time gate that spends extra compute only when the VLA policy becomes uncertain. CAPS keeps the base policy unchanged, computes contextual SNR from action entropy, and runs Metropolis-Hastings search over future action chunks when entropy crosses a threshold. When uncertainty stays low, it uses greedy execution.

This is a fit for long-horizon tasks where a single locally plausible action can lose the instruction goal. The validation path is direct: log entropy during rollouts, replay high-entropy windows with action-chunk search, and compare task success and latency against the base policy. CAPS reports 47.4% average success on RoboTwin 1.0 with π0, compared with 32.2% for π0 and 41.3% for π0 plus TACO. On Simpler-WindowX, it reports 60.5% average success, ahead of π0 at 48.0% and π0 plus TACO at 55.5%.

Retail robotics groups can start with in-store human activity capture before expensive robot teleoperation. SABER shows a concrete data workflow: record natural grocery-store work with egocentric and fixed 360° cameras, estimate hand and body motion, manually correct pose estimates, and retarget the result into robot-compatible action targets. The resulting streams cover latent-action sequences, dexterous hand-pose trajectories, and whole-body motion for a humanoid.

This directly addresses the adoption blocker in store robots: general VLA training data often misses shelf picking, fridge opening, basket loading, floor retrieval, occlusion, lighting, and varied packaging. SABER reports 44.8K training samples from about 100 hours of real store capture and post-trains GR00T N1.6 to 29.3% mean success across 10 RoboBenchMart tasks, compared with 13.4% for simulation-only fine-tuning. A practical pilot would capture one aisle, convert a small set of worker actions into the same three supervision streams, and test whether task success improves on the store-specific actions that fail most often.