Scaling expert demonstrations alone does not enable generalist robot policies to improve from deployment. Real-world rollouts contain successes, suboptimal behavior, and failures: imitating every recorded action can reinforce poor behavior, while filtering out failures discards evidence of physical interactions and their consequences. We introduce LoopWAM, a closed-loop world-action model that learns directly from this full quality spectrum, using rollouts from both itself and other deployed policies. Source-dependent noise sampling unifies behavior learning and action-conditioned world modeling, while quality steering and quality-weighted supervision retain full video supervision across qualities and regulate action imitation. Video and action tokens interact throughout parallel denoising, allowing learned consequences to inform action generation directly. On RoboChallenge Table30-V2, three deployment-and-training rounds raise success from 23.67% to 43.33%, a 19.7-percentage-point gain, with a final score of 56.80. LoopWAM also achieves 94.5% average success on RoboTwin~2.0 and 49.5% overall success on RoboCasa365 without additional embodied pretraining, obtaining the best overall results among the compared methods across all three benchmarks. On challenging shirt folding task, two rounds of deployment-driven improvement after single-task post-training raise success from 13.20% to 55.56%. These results demonstrate closed-loop improvement from mixed-quality experience under both multi-task learning and single-task post-training. orange{Project Page}
Figures & tables
Figure 1: LoopWAM learns from deployment. Mixed-quality rollouts return to training through quality steering, source-dependent noise sampling, and quality-weighted supervision. Center: closed-loop self-improvement on RoboChallenge Table30-V2, where three rounds of deployment and retraining increase overall success from 23.67% to 43.33%. Bottom: benchmark comparisons.
Figure 2: Overview of LoopWAM. LoopWAM is structured as a Mixture-of-Transformer (MoT) architecture, where bidirectional attention enables feature interaction between the video and action branches. Training combines quality-conditioned instructions, source-dependent noise sampling, and quality-weighted action supervision. Expert trajectories supervise both modalities, while lower-quality rollouts retain full world supervision and contribute to action imitation according to their quality. The trained model is deployed using the expert score q⋆ , and recorded experience returns to the next training round.
Method
ALOHA
DOS-W1
ARX5
UR5
Overall
Score
SR
Score
SR
Score
SR
Score
SR
Score
SR
debuggers_vla
5.40
0.00
5.15
2.00
53.21
34.29
27.83
23.33
18.72
11.00
W0
12.25
2.00
4.40
0.00
48.64
30.00
49.50
43.33
21.85
12.00
test_01
6.25
1.00
8.95
3.00
49.50
28.57
61.17
43.33
22.73
12.33
holobrain
11.75
3.00
13.30
6.00
53.64
30.00
70.67
43.33
27.93
14.33
Baseline Model
21.55
6.00
17.35
6.00
60.14
32.86
44.83
26.67
31.48
14.33
Table 1: Performance on RoboChallenge Table30-V2. We report per-embodiment success rates (SR, %) and overall performance across four robot embodiments. Baselines are sorted by overall SR. Bold denotes the best performance, while blue and light blue highlight LoopWAM and the runner-up, respectively.
Method
Embodied PT.
Clean
Rand.
Average
π0 ( Black et al., 2024 )
✓
65.9
58.4
62.2
X-VLA ( Zheng et al., 2025 )
✓
72.9
72.8
72.9
π0.5 ( Physical Intelligence et al., 2025 )
✓
82.7
76.8
79.8
ABot-M0 ( Yang et al., 2026 )
✓
81.2
80.4
80.8
Motus (from Wan2.2) ( Bi et al., 2025 )
✗
77.6
77.0
77.3
Motus ( Bi et al., 2025 )
✓
88.7
87.0
87.8
Table 2: Performance on RoboTwin 2.0. We report success rates (SR, %) under clean and randomized evaluation settings. Embodied PT. denotes additional embodied-data pretraining before benchmark training. LoopWAM is trained with the mixed expert–rollout corpus described in Section 4.2 . Blue and light blue highlight LoopWAM and the runner-up, respectively.
Method
Embodied PT.
Atomic Seen
Composite Seen
Composite Unseen
Overall
Diffusion Policy ( Chi et al., 2025 )
✗
15.7
0.2
1.3
6.1
π0 ( Black et al., 2024 )
✓
34.6
6.1
1.1
14.8
π0.5 ( Physical Intelligence et al., 2025 )
✓
39.6
7.1
1.2
16.9
GigaWorld-Policy 0.1 ( Ye et al., 2026a )
✓
44.4
11.8
2.9
20.7
GR00T N1.6 ( NVIDIA et al., 2025 )
✓
51.1
9.4
1.7
21.9
GR00T N1.5 ( NVIDIA et al., 2025 )
✓
50.7
14.8
2.7
23.9
Table 3: Performance on RoboCasa365. We report success rates (SR, %) across three evaluation splits. Baseline results are obtained from the official leaderboard. Embodied PT. follows the definition in Table 2 . Blue and light blue highlight LoopWAM and the runner-up, respectively.
Figure 3: Closed-loop improvement from real-robot deployment. Top: three deployment-and-training iterations on RoboChallenge Table30-V2, covering 30 tasks across four robot embodiments. Bottom: two deployment-and-training iterations on shirt folding after task-specific expert post-training of the DOS-W1 model. Each plot reports success rate and score over successive improvement rounds.
Steering
Act.
Clean
Rand.
Avg.
None
✗
73.5
71.1
72.3
Extra-coarse
✗
92.2
93.3
92.8
Extra-coarse
✓
93.9
93.3
93.6
Coarse
✗
93.2
92.2
92.7
Coarse (ours)
✓
94.6
94.4
94.5
Table 4: Ablation study on RoboTwin 2.0. We evaluate quality steering with different annotation schemes and assess the effect of quality-weighted rollout action supervision (Act.).
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Robot
Ordered subtask instructions
Press the button
ARX5
1. Press the pink button. 2. Press the blue button. 3. Press the green button. 4. Press the yellow button, then retract the gripper.
Water the flowers
ARX5
1. Pick up the watering can and water the potted plant. 2. The watering is finished, put the watering can down.
Shred paper
UR5
1. Put the paper into the shredder. 2. Retract gripper.
Stamp positioning
ALOHA
1. Pick up the stamp. 2. Dip into the ink. 3. Stamp the signature area and return.
Appendix
Table 5: Table30-V2 subtask instructions, listed in execution order.
Figure 4: Button pressing (ARX5). Subtask segmentation example. The initial observation (Stage 1) and subsequent stage observations are shown in order, each paired with its subtask instruction.
Figure 5: Plant watering (ARX5). Subtask segmentation example. The initial observation (Stage 1) and subsequent stage observations are shown in order, each paired with its subtask instruction.
Figure 6: Paper shredding (UR5). Subtask segmentation example. The initial observation (Stage 1) and subsequent stage observations are shown in order, each paired with its subtask instruction.
Figure 7: Stamp positioning (ALOHA). Subtask segmentation example. The initial observation (Stage 1) and subsequent stage observations are shown in order, each paired with its subtask instruction.
Figure 8: RoboChallenge Table30-V2 leaderboard. LoopWAM ranks first with an overall score of 56.80 and a success rate of 43.33%, using a multi-task policy for each embodiment.
Figure 9: Qualitative examples of iterative execution and failure recovery. The first attempt may fail due to execution uncertainty, incomplete interaction, or object displacement. Subsequent attempts generate corrective behaviors and achieve successful task completion.
Figure 10: Qualitative comparison between predicted future observations from LoopWAM and ground-truth robot executions. The model captures the temporal evolution of robot-object interactions and predicts plausible future visual states.
Figure 11: Representative tasks on the ARX robot platform. The examples illustrate diverse manipulation scenarios used for evaluating LoopWAM across different robot embodiments.
Figure 12: Representative tasks on the ALOHA robot platform. The examples include coordinated manipulation and object interaction requiring temporally consistent control.
Figure 13: Representative tasks on the UR5 robot platform. The examples demonstrate single-arm manipulation in structured environments.
Figure 14: Representative long-horizon tasks on the W1 robot platform. The examples demonstrate sequential manipulation behaviors in complex environments.
World Action Models (WAMs) emerge as a promising foundation for policy refinement by predicting the consequences of sampled actions. However, visually plausible predictions can mislead policy refinement if they fail to reflect the input actions. To address this mismatch, we introduce R2-WAM, a two-stage repair-and-reject post-training framework that first improves the consistency of predicted futures with input actions, then uses these futures to select inferior action samples for negative fine-tuning. The repair stage grounds imagination in observed robot behavior through a kinematic alignment score that measures agreement between predicted and demonstrated motion, enabling the predicted video to faithfully reflect its input actions. Using the repaired video model, the rejection stage compares imagined outcomes of sampled and demonstrated actions, selectively applying negative fine-tuning to samples whose predicted task progress falls below the demonstrated reference by a prescribed margin. Together, the two stages extend video prediction from representation learning to consequence-based policy refinement without additional environment interaction or changes to the inference procedure. R2-WAM achieves 93.8% average success on RoboTwin 2.0 across clean and randomized settings. On the long-horizon real-world Fold Shirt task, it achieves 87.5% average success, compared with 0% for Fast-WAM. Our project page is available at https://r2-wam.github.io/.
Ruiyan Xu, Haisheng Su, Sixu Lin +4
The Chinese University of Hong Kong, Shenzhen · Manifold AI · Southeast University +1
Improving robot policies on new tasks without collecting additional expert demonstrations remains a central challenge in robot learning. World action models (WAMs) use broad video priors to jointly predict future videos and actions, offering a potential source of supervision for adapting to new tasks. However, generated videos may fail to depict task completion, and even visually successful videos may be paired with inconsistent actions that lead to execution failure. We propose EVO-WAM, a framework that adapts WAMs to unseen tasks by learning from their own generated video-action trajectories, without executing candidate actions in an external environment. First, we augment WAM training with state prediction and anchored multi-frame context to enable complete autoregressive rollouts without external execution feedback. Second, we identify reliable training experience by selecting task-completing prefixes with a vision-language model and verifying their video-action consistency with an inverse dynamics model. Third, we iteratively train the WAM on verified prefixes and generate new rollouts with the updated model. On seven unseen RoboTwin 2.0 tasks, EVO-WAM increases average success rates from 26.9% to 68.0% for Cosmos3 and from 28.5% to 46.4% for DreamZero, reaching approximately 2.5× and 1.6× their initial success rates. On three unseen long-horizon composite tasks in the real world, it improves Cosmos3's average success rate from 20.0% to 76.7%, a gain of 56.7 percentage points. Project Page: https://evo-wam.github.io/.
World Action Models (WAMs) generate both future video and robot actions, offering two connected outputs for post-training supervision. How can a pretrained WAM learn from a stronger Teacher on the histories it encounters during execution? We present WAM-OPD, which collects Student rollout histories and queries a Teacher for paired video and action targets. The Student learns from both targets while retaining its one-step video and action generation at deployment. Across 12 RoboTwin 2.0 tasks, WAM-OPD improves average success from 33.8% to 65.7%; across four real-robot tasks, it improves average success from 51.4% to 64.6%. With the collected Student histories held fixed, joint video-action supervision achieves the highest observed success on all three ablation tasks, while either modality alone also improves performance. A separate comparison with Teacher-generated histories finds task-dependent differences between the two history sources. These results demonstrate the value of paired video-action supervision for improving WAM policies without increasing their deployed sampling budget.