EVO-WAM: Robots Self-Train on Imagined Trajectories, Lifting Unseen Task Success from 27% to 68%
量子位 · wechat · 2026-10-19
The EVO-WAM team introduces a self-improvement loop for World Action Models that requires no new expert demos or environment trial-and-error: the model generates full candidate video-action trajectories from a single initial observation, a VLM verifies task completion, and an Inverse Dynamics Model checks that generated actions match the video. Only validated clips are mixed back into training data for iterative fine-tuning.
Results: on seven unseen RoboTwin2.0 sim tasks, Cosmos3's average success rate rose from 26.9% to 68.0% over four rounds (DreamZero: 28.5%→46.4%); on three unseen real-robot long-horizon tasks, from 20.0% to 76.7%, with prior-task performance largely retained. Ablations show VLM-only filtering reaches just 43.7% versus 68.0% with IDM action-consistency checks. Paper and code are open-sourced.
More from Embodied
- MIT CSAIL's X-Hinges 3D-Prints Everyday Objects With Built-In Sensing Across Three Motion Axes — MIT News AI · 2026-10-23
- Formic raises $50M Series B to scale physical AI across 100+ US factories — garrytan · 2026-10-23
- fal starts fine-tuning a physical robot to draw, results next week — OdinLovis · 2026-10-23
- PEARL demo: an LLM agent that checks experiment results, revises plans, and re-tests — KevinKaichuang · 2026-10-23
- CU Boulder's ATLAS Institute Opens Tenure-Track Robotics Professorship — ryosuzk · 2026-10-23
- MiTaS multi-resolution tactile imitation learning hits 80% success vs 35.5% vision-only baseline — Jan_R_Peters · 2026-10-23