Tsinghua-affiliated LoopWAM lifts robot success rate from 23.7% to 43.3% via self-deployment data
jiqizhixin · x · 2026-10-30
On RoboChallenge Table30 V2, LoopWAM—by Tsinghua University researchers and collaborators—raised task success from 23.67% to 43.33% over three deployment-driven improvement rounds, without extra expert demonstrations. The method assigns different roles to mixed-quality data: video prediction learns from the full rollout pool including failures, action training is weighted by trajectory quality, and quality conditioning steers inference toward better behavior, closing a deploy-train loop that reuses otherwise discarded experience.
More from Embodied
- Toyota bZ4X with Waymo self-driving kit shown in official render, rollout teased — reed · 2026-10-30
- Hacker builds a trick-or-treat robot with RealSense D457 and Jetson AGX Orin — chrismatthieu · 2026-10-30
- Machenta's instant robot learning now runs on SO-101 robots, community role open — Ed__Johns · 2026-10-30
- Oopsie-data: Building the first open dataset of robot failure trajectories — chris_j_paxton · 2026-10-30
- Podcast: Are we ready for brain-computer implants? Paradromics CEO weighs in — GregCook2011 · 2026-10-30
- Open NV-Segment-CT segments all 74 abdominal CT structures in 13 seconds on one GPU — MaziyarPanahi · 2026-10-30