CoachSI post-training of OpenWAM on RoboCasa365
Researchers at the University of Illinois Chicago collect rollouts from the task start. Frozen Qwen3.8-27B chooses an instruction at each OpenWAM action chunk. The researchers relabel successful trajectories with the original task instruction and fine-tune OpenWAM. Evaluation uses separate kitchen scenes and only the task instruction.
This announcement
Across 18 long-horizon tasks, 20 unseen scenes and three trials per scene give 1,080 episodes. Unassisted success rises from 47.4% to 56.8% with image-only coaching at collection or 59.3% with simulator-state coaching. Training uses 3,000 steps, batch size 16, 80% human demonstrations and 20% comparison data. Only video-backbone self-attention is updated, and the action backbone stays frozen. OpenWAM scores 79.2% on atomic tasks. The human-demonstration and self-collected-success baselines reach 49.4% and 51.9%. The variable 20% of the data has a matched budget of 530 samples per task and comes from extra human demonstrations, self-collected successes or coached successes. The image-only coach receives three camera views. Oracle next-step prompts yield 4.2 times the original instruction’s subgoal completion rate within 320 steps. The two base policies contribute a total of 500 stalled states from 15 tasks. A stall means 320 steps without subgoal progress. Each restored state has six continuations per instruction source with matched seeds. Subgoal completion is measured within 320 steps and task success within 640.
The paper identifies a public OpenWAM checkpoint and gives no exact version, checkpoint identifier or revision. Training code is marked coming soon. Derived weights and their license, as well as the code license, remain unconfirmed.



