LoopWAM gains 19.7 points on real robots by retraining on its own labeled rollouts

LoopWAM is a closed-loop world-action model from Tsinghua University, Dexmal and Loopeva, with 14 authors and Xiuwei Xu, Meng Zhang and Jiwen Lu as corresponding authors. The team posted it to arXiv on Oct 13, 2026 and revised it on Oct 15, with code on GitHub. On RoboTwin 2.0's 50 bimanual tasks it reports 94.5% average success with no additional embodied pretraining, ahead of LingBot-VA 2.0 at 93.6% and π0.5 at 79.8% in the same table.

Mixed-quality rollouts, failures included, pay off only once they are labeled, and crude tags do the job: the same RoboTwin data scores 72.3% without quality steering and 92.8% with three-level tags, and quality-weighted action supervision on the rollouts lifts that to 93.6%. The real-robot crown is thin by comparison, 2.67 points over Dexmal's own DM0.5 on a benchmark Dexmal helped launch and 0.33 points against the 43.0% Dexmal publishes for DM0.5, and it appears only after three deployment rounds, from a model that started 17 points behind.

Failed runs train the world model, not the arm

Every trajectory in the training set gets a numeric quality score attached to its instruction text, written as "Quality: q", and at deployment the model is conditioned on the expert score to steer toward expert behavior. Video prediction is supervised on all trajectories, failures included. Action imitation learning is weighted by quality: expert trajectories get full action supervision, and failed ones get none, so their actions are never copied while they still train the world model to predict what happens after an action. The authors denoise video and action tokens in one parallel process, so predicted consequences feed directly into action generation, and a "source-dependent noise sampling" step lets recorded rollout actions guide prediction of their observed consequences.

The ablation shows how crude the labels can be. With no quality steering the RoboTwin average is 72.3%. Extra-coarse markers, where every failed rollout is low, every success medium and every expert high, reach 92.8% without rollout action supervision and 93.6% with it. One numeric score per policy checkpoint, with failures at the minimum, reaches 92.7% without and 94.5% with, the full method. The authors report that the coarse scheme needs a single human annotation per checkpoint, and conclude that fine-grained annotation is not required to benefit from steering, which they call the key ingredient.

Three rollout rounds add 19.7 points, and the biggest jump is the second

The hardware stage is RoboChallenge's Table30-V2, a real-robot benchmark of 30 tabletop manipulation tasks across four embodiments, ALOHA, DOS-W1, ARX5 and UR5, whose technical report is by the RoboChallenge Dexmal Team. The authors trained one multi-task LoopWAM per embodiment on expert demonstrations, then ran three deployment rounds. Each round collected about 100 rollout trajectories per task, annotated them by quality, and retrained on the expert data plus every earlier rollout.

Overall success rose 23.67% to 28.00% to 37.67% to 43.33%, a gain of 19.7 percentage points, while the overall score rose 36.10 to 42.63 to 51.65 to 56.80. The round gains in success were 4.33, 9.67 and 5.66 points, so the second round carried the most. On the authors' own procedure, 100 rollouts per task across 30 tasks is roughly 3,000 trajectories per round, or about 9,000 over the three rounds.

The Table30-V2 lead is 2.67 points, or 0.33 by Dexmal's own count

LoopWAM ranks first in both success and score among the leaderboard entries the paper compares, with runner-up VLA-DM0.5 at 40.67% success and 54.42 score, a lead of 2.67 success points at the snapshot accessed Sept 22, 2026. Dexmal's own page for DM0.5, a VLA with a Gemma3 4B vision-language backbone and a 680M action expert, reports 43.0% success and the same 54.42 score, and against that published figure the LoopWAM lead falls to 0.33 points.

By embodiment the result splits. LoopWAM finishes ahead on ALOHA, 26.00% to 18.00%, and on UR5, 76.67% to 70.00%, and ties DM0.5 on DOS-W1 at 42.00%. On ARX5 it loses, 55.71% to 58.57%.

The strongest case against the ranking is the starting line. Before any deployment round, the expert-trained LoopWAM stood at 23.67%, 17 points behind DM0.5's 40.67%, and it needed all three rounds to pass it. Dexmal, founded in March 2025 and based in Beijing, co-launched RoboChallenge, and DOS-W1 is its own open-source hardware.

The same recipe accepts rollouts from other deployed policies, even poorly performing ones, so one generalist policy can train on another's mistakes. On RoboTwin 2.0 it trained on 22,500 rollout trajectories from itself, X-VLA and π0.5 across 50 tasks and three policies; on RoboCasa365 it used rollouts from itself and RLDX-1.

Physical Intelligence's approach is the closest reference point. The authors write that π0.6 uses advantage-conditioned learning to fold in autonomous experience and interventions, and that π0.7 uses quality, speed and mistake metadata to absorb mixed-quality data, and they argue those methods carry no explicit objective for the physical dynamics in action-video transitions, so a failed rollout there gives a behavior target without teaching the model the consequences of that behavior. Pi's API beta, announced the same day as the preprint, describes a loop of collecting demonstrations, training a model, testing it on the robot and using the failures to decide what data to collect next.

Zero-shot transfer is where the paper is weakest. On RoboCasa365's 16 Composite-Unseen tasks, held out for zero-shot evaluation, LoopWAM scored 14.9%, below PRTS's 18.8% in the same table, although it still led the benchmark overall at 49.5% against ABot-M0.6's 46.6%. If that 14.9% on held-out simulated tasks carries over to new kitchens, the world model's advantage is tied to setups it has already seen.

Shirt folding is the one task that shows the loop end to end. On DOS-W1, after single-task post-training the policy folded shirts at 13.20% success, and two deployment rounds raised that to 32.54% and then 55.56%, with the score going 14.44 to 38.89 to 67.78. That is more than four times the starting success rate on a deformable-object task that takes multiple steps.