LoopWAM

Closed-loop generalist policy improvement from deployment rollouts.

Project page · Code

LoopWAM turns real-world experience into better robot policies. Expert demonstrations sit alongside successful, suboptimal, and failed rollouts. Every rollout teaches the world model. Action imitation is weighted by quality, and failed actions are not copied.

Method

Video and action tokens interact through bidirectional attention while they are denoised together.

• Joint video–action modeling. The two streams attend to each other at every layer.
• Source-dependent noise. Lower-noise rollout actions guide video prediction toward the physical consequences that were observed.
• Quality-aware learning. Quality conditions and weights guide action learning. Every rollout teaches video prediction, without imitating failed actions. At deployment, an expert-quality target steers the policy.

Leaderboards

Best overall performance among the methods compared in the paper.

Benchmark	LoopWAM
RoboChallenge Table30-V2	43.33% success · 56.80 score
RoboTwin 2.0	94.5% average
RoboCasa365	49.5% overall

RoboChallenge Table30-V2

Real-world, 30 tasks, 4 embodiments. One multi-task model per embodiment. Overall score uses task weights 10 / 10 / 7 / 3.

Method	ALOHA	DOS-W1	ARX5	UR5	Score	SR (%)
LoopWAM	26.00	42.00	55.71	76.67	56.80	43.33
VLA-DM0.5	18.00	42.00	58.57	70.00	54.42	40.67
my16	13.00	27.00	54.29	46.67	41.27	30.67
LoopWAM (expert-trained)	8.00	38.00	24.29	26.67	36.10	23.67
JV0	5.00	16.00	20.00	43.33	27.58	16.00
mc	7.00	0.00	47.14	20.00	23.63	15.33
MagicBot	0.00	5.00	32.86	53.33	20.82	14.67
holobrain	3.00	6.00	30.00	43.33	27.93	14.33
Baseline Model	6.00	6.00	32.86	26.67	31.48	14.33
test_01	1.00	3.00	28.57	43.33	22.73	12.33
W0	2.00	0.00	30.00	43.33	21.85	12.00
debuggers_vla	0.00	2.00	34.29	23.33	18.72	11.00

Columns ALOHA, DOS-W1, ARX5, and UR5 are success rates (%).

RoboTwin 2.0

50 bimanual tasks. No additional embodied pretraining. Clean 94.6%, randomized 94.4%.

Method	Embodied pretraining	Clean	Randomized	Average
LoopWAM	No	94.6	94.4	94.5
LingBot-VA 2.0	Yes	93.8	93.4	93.6
AHA-WAM	No	93.4	92.2	92.8
LingBot-VA	Yes	92.9	91.5	92.2
Fast-WAM	No	91.9	91.8	91.8
Motus	Yes	88.7	87.0	87.8
ABot-M0	Yes	81.2	80.4	80.8
π0.5	Yes	82.7	76.8	79.8
Motus (from Wan2.2)	No	77.6	77.0	77.3
X-VLA	Yes	72.9	72.8	72.9
π0	Yes	65.9	58.4	62.2

RoboCasa365

50 evaluation tasks. Overall success uses weights 18 / 16 / 16. All 16 Composite-Unseen tasks are held out of training.

Method	Embodied pretraining	Atomic seen	Composite seen	Composite unseen	Overall
LoopWAM	No	79.7	50.3	14.9	49.5
ABot-M0.6	Yes	79.4	48.3	7.9	46.6
ABot-M0.5	Yes	75.6	37.7	3.3	40.3
PRTS	Yes	66.3	30.3	18.8	39.6
RLDX-1	Yes	67.6	27.9	8.5	36.0
GR00T N1.5	Yes	50.7	14.8	2.7	23.9
GR00T N1.6	Yes	51.1	9.4	1.7	21.9
GigaWorld-Policy 0.1	Yes	44.4	11.8	2.9	20.7
π0.5	Yes	39.6	7.1	1.2	16.9
π0	Yes	34.6	6.1	1.1	14.8
Diffusion Policy	No	15.7	0.2	1.3	6.1

Multi-task self-improvement

30 tasks, four embodiments, three rounds of real-world deployment. The clip below is 30 full rollouts, sped up to 30 seconds.

Stage	Success	Score
Expert-trained	23.67%	36.10
Round 1	28.00%	42.63
Round 2	37.67%	51.65
Round 3	43.33%	56.80

Tasks

Click a still to open the rollout video.

ALOHA

ARX5

UR5

DOS-W1

Shirt folding

After expert post-training, two deployment rounds improve DOS-W1 shirt folding from 13.20% to 55.56%.

Stage	Training data	Success	Selected trials
Stage 1	Expert demonstrations	13.20%	1 success, 8 failures
Stage 2	Expert + round 1 rollouts	32.54%	3 successes, 6 failures
Stage 3	Expert + rounds 1 and 2	55.56%	5 successes, 4 failures

Stage 1, the success:

Stage 2 successes:

Stage 3 successes:

Action and video generation

Each pair is the observed video on the left and LoopWAM’s prediction on the right.

Arrange Fruits

Pen Container

Stack Bowls

Lint Roller

Checkpoints

Each directory contains one weight file and its dataset_stats.json.

Benchmark	Directory	Checkpoint
Table30-V2, UR5	table30v2_ur5/	step_029925.pt
Table30-V2, ALOHA	table30v2_aloha/	step_061320.pt
Table30-V2, ARX5	table30v2_arx5/	step_057355.pt
Table30-V2, DOS-W1	table30v2_w1/	step_072970.pt
RoboTwin 2.0	robotwin/	step_037645.pt