Reliable Robot Manipulation Systems

Teams deploying a pretrained VLA can wrap it in a planner that records completed stages, contact outcomes, and retry history. The VLA handles short contact-rich actions; explicit primitives handle grounding, free-space motion, re-staging, and release. Harness VLA used this structure to raise LIBERO-Pro success from 50.0% for the direct frozen baseline to 82.4% without VLA fine-tuning. TFP also shows that episode-local task state can resolve visually similar scenes that require different actions: real-robot object-swap success increased from 3/20 to 15/20.

A practical trial can start with one multi-stage task that currently fails after an empty grasp, occlusion, or object swap. Log a compact state containing the current stage, elapsed time, last contact result, and retry count. Compare the wrapped policy with the unchanged controller across layout changes, measuring full-task success, repeated-action errors, retries, and added latency. This test will show whether execution state is a cheaper reliability gain than collecting another task-specific fine-tuning set.

Robot-learning teams can add a relabeling pass to their post-training pipeline so coherent failures produce usable supervision. A VLM watches a failed rollout, describes the behavior that actually occurred, and scores the rollout against that new instruction. Learning from Hindsight kept roughly 70% to 80% of rollout groups usable, reached standard GRPO performance in about five training steps compared with nearly 30, and achieved 56% real-robot success after 160 rollouts versus 22% for GRPO.

The cheapest check is an offline audit of a recent failed-rollout batch. Sample 100 failures, generate hindsight instructions, and have operators review whether each description is correct, specific, and safe to reinforce. If agreement is high, run a short post-training comparison using the same rollout budget and track usable-group rate, original-task success, and any increase in unintended behaviors. This workflow is most relevant where physical collection time and sparse rewards dominate training cost.

Policy evaluation for contact-rich manipulation should include execution time, force exceedances, and low-level tracking error alongside task success. B-spline Policy shows that continuous action curves can be retimed and sampled at high control frequency: table-cleaning time fell from 23.57 to 11.80 seconds while success moved from 13/20 to 14/20. Its 4× Speed Stacking result also fell to 0/20, exposing a controller-speed boundary. PAC-ACT reports a 46-fold reduction in force readings above 60 N after chunk-level reinforcement-learning post-training.

A useful test bench can replay the same demonstrations through discrete action chunks and continuous curves at several speed multipliers, then record success, completion time, peak force, time above the force limit, and joint-tracking error. Include at least one precision-contact task and one long-horizon task. DexVerse provides a warning against relying on easy pick-and-place checks: the top mean success across its 19-task subset was 0.34, and every tested policy scored zero on PushT. The resulting speed-force envelope gives deployment engineers an explicit operating limit for each controller and task.