Event

Proposal of the SimPT automated simulation post-training framework

1 reports · 1 sources

Researchers including Caicheng Wang published the SimPT framework paper on arXiv, integrating policy rollout, corrective data collection, policy training, and independent evaluation to automate human-in-the-loop post-training of vision-language-action (VLA) models. The paper reports: in real-machine HIL post-training, the success rate on the Ethernet cable insertion task across 25 scenarios rose from 72% to 96%; its Progress Graph reduced recovery attempts by 75.0% relative to a fixed-interval snapshot baseline, and the recovery success rate rose from 93.4% to 98.9%. The researchers also compared four post-training methods—RECAP, HG-DAgger, DPO+SFT, and FlowPRO—on seven manipulation tasks in RoboTwin, where one improvement raised the overall success rate on held-out simulation trials from 76.50% to 79.36%.