World action models (WAMs) have shown promising generalization and data efficiency in robot manipulation by coupling action generation with predictions of future visual dynamics. Translating these capabilities into reliable task execution, however, remains a challenge for existing post-training methods. Offline supervised fine-tuning provides dense supervision on expert trajectories, but its flow-matching action expert is trained only on noised demonstration actions and never on the intermediate states produced by its own sampler; task-specific adaptation can also compromise pretrained capabilities. On-policy reinforcement learning trains on the policy's own behavior, yet sparse task rewards make exploration and credit assignment difficult, while dense reward shaping still provides scalar evaluations rather than direct action targets. We introduce HINT-WAM, the first on-policy self-distillation framework that supervises the action expert on its own sampling trajectories using hindsight, the future video of the demonstration, as privileged information available to the WAM itself. For each demonstration observation, the student generates its action through its own denoising trajectory, and at the states along this trajectory the same model acts as a teacher conditioned additionally on this hindsight, providing a conditional action distribution as a dense training target, while the student sees only deployment-time inputs and needs no environmental rewards or simulator interaction. To transfer this supervision to a flow-matching action expert, HINT-WAM distills the teacher into the student with an energy score, a distribution-level objective that matches the teacher's action distribution from the student's own samples rather than regressing its velocity field. The project web is available in https://tuandattt.github.io/hintwam/.
Figures & tables
Figure 2: Overview of HINT-WAM. Stage 0 (top left): the action expert is fine-tuned with flow matching on demonstrations, conditioned on the key/value cache of the frozen video expert, giving the student initialization. Stage A (top right): with the SFT model frozen, a hindsight encoder compresses the demonstration’s future observations into hindsight tokens, which the action expert reads through gated cross-attention adapters; only the encoder and adapters are trained. Stage B (bottom): the student integrates its own ODE without gradients, and its implied actions at selected denoising points are matched to a single sample from the frozen hindsight teacher with the energy score, together with a behavior-cloning anchor. Only the student is deployed.
Method
LIBERO-90
RoboTwin 2.0
Clean
Random.
Fast-WAM [ 45 ]
15.56
86.48
83.44
SFT
78.88
–
–
Full SFT
81.24
84.12
83.24
BC control
71.56
85.36
82.08
Self-imitation
79.62
–
–
Table 1: Post-training performance on LIBERO-90 and RoboTwin 2.0. All values are success rates (%). –: not applicable or not evaluated.
Method
New tasks
Original suites
LIBERO-90
Spatial
Object
Goal
Long
Avg
Fast-WAM [ 45 ]
15.56
96.8
99.6
95.8
94.0
96.55
Without replay
SFT
78.88
58.6
33.2
53.4
4.6
37.45
Full SFT
81.24
54.4
30.2
44.6
2.2
32.85
BC control
71.56
55.4
25.6
47.0
4.0
33.00
Table 2: Continual learning on LIBERO. Success rates (%) on the new LIBERO-90 tasks and the four original suites after post-training on LIBERO-90, without and with experience replay.
Figure 4Figure 5
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
HINT-WAM
Baselines
Stage 0 (SFT)
Stage A (teacher)
Stage B
BC control
Full SFT
Init.
released
SFT@3k, frozen
SFT@3k
SFT@3k
released
Trainable
action expert
ψ (enc.+adapt.)
action expert
action expert
video + action experts
#Params
1.02B
13M
1.02B
1.02B
6.02B
Loss
Eq. 1 , λBC=2
Eq. 6
Eq. 9
Eq. 1 , λBC=1
Eq. 1 , λBC=2
Learning rate
10−4
3×10−4
10−4
10−4
10−4
Appendix
Table 3: Training settings of each stage and baseline on LIBERO-90. All runs share the optimizer, schedule, and effective batch size of 192.
Stage
Steps
s/step
Hours
GPU-h
SFT (action expert)
3k / 5k
40.2
33.5 / 55.8
268 / 447
Stage A (teacher)
2k
7.9
4.4
35
Stage B (distill.)
2k
60.6
33.7
269
BC control
2k
40.2
22.3
179
Full SFT
5k
91.4
127
1,016
Pipeline
Hours
GPU-h
Success
Appendix
Table 4: Training cost on LIBERO-90. Top: individual stages. Bottom: complete pipelines and their LIBERO-90 success rates.
Teacher
Loss change
LIBERO-90, built on SFT
−3.28% ( t=5.77 )
LIBERO-90 + replay, built on SFT + replay
−3.06% ( t=4.68 )
RoboTwin 2.0, built on released checkpoint
−5.37% ( t=3.43 )
LIBERO-90, 1,000 vs. 2,000 steps
−2.36%→−3.28%
Appendix
Table 5: Hindsight teacher checks. Relative change in the held-out action loss of the teacher with versus without hindsight ϕ .
Comparison
Mean diff.
SE
t
Wins / losses / ties
Sign test p
HINT-WAM vs. SFT
+6.30
1.76
3.58
36 / 22 / 32
0.087
HINT-WAM vs. self-imitation
+5.56
1.88
2.96
47 / 23 / 20
0.0056
HINT-WAM vs. shuffled hindsight
+4.90
1.78
2.76
45 / 25 / 20
0.022
HINT-WAM vs. velocity matching
+6.16
1.59
3.62
48 / 17 / 25
0.00042
Velocity matching vs. SFT
+0.14
1.52
0.09
30 / 37 / 23
0.46
HINT-WAM + replay vs. SFT + replay
+4.86
–
2.66
44 / 23 / 23
0.014
Appendix
Table 6: Paired statistics on LIBERO-90. Differences are in % over 90 tasks with 50 trials each.
Figure 8: Per-task gain of HINT-WAM over SFT on LIBERO-90. (a) Gain on each task, sorted. (b) Mean gain ( ± standard error) with tasks grouped by the success rate of SFT.
HINT-WAM vs. SFT
MSE vs. SFT
HINT-WAM vs. MSE
Task
SFT
Ours
Task
SFT
MSE
Task
MSE
Ours
52
80
58
74
80
32
76
86
60
53
60
40
89
90
44
53
52
30
74
80
60
86
60
28
75
48
30
5
100
86
58
80
52
21
96
80
78
100
86
6
100
74
30
94
80
Appendix
Table 7: Largest per-task drops on LIBERO-90. Success rates (%) over 50 trials per task.
Comparison
Clean
Randomized
HINT-WAM vs. BC control
+1.68 (0.84, t=2.00 ), 12 / 9 / 4
+1.60 (1.32, t=1.21 ), 14 / 9 / 2
HINT-WAM vs. released
+0.56 (1.19, t=0.47 ), 11 / 9 / 5
+0.24 (1.04, t=0.23 ), 13 / 9 / 3
BC control vs. released
−1.12 (1.08, t=−1.04 ), 9 / 11 / 5
−1.36 (1.43, t=−0.95 ), 8 / 15 / 2
Appendix
Table 8: Paired tests on the RoboTwin 2.0 hard subset. Mean difference in % (SE, t ) and wins / losses / ties over 25 tasks.
Figure 9: Toy example of distillation under hidden information. A hidden variable c selects one of two correct actions; the teacher observes it, the student does not. (a) Teacher samples. (b) Squared-error distillation collapses to the mean of the two modes. (c) The energy score recovers both.
Figure 10: Rollouts on RoboTwin 2.0 ( stack_bowls_three , clean setting). SFT (top row of each panel) fails, while HINT-WAM (bottom row) stacks all three bowls from the same initial state.
World Action Models (WAMs) improve robot policy learning by jointly modeling future visual dynamics and actions. However, their scalability and generalization remain constrained by their reliance on costly expert demonstrations. We challenge this by asking whether future supervision for WAMs must originate from target-task expert trajectories. In this paper, we propose Vid2WAM, an offline distillation framework that transfers visual diffusion priors from a large video foundation model into a compact WAM student. Given an observation and language instruction, Vid2WAM distills supervision through two complementary channels: task-conditioned future rollouts directly supervise the student's future prediction branch, while an inverse dynamics model recovers embodiment-specific pseudo-actions for action learning. To robustly integrate synthetic and real supervision, we introduce source-aware residual action adaptation that learns source-specific corrections around a shared action backbone and mitigates interference from noisy pseudo-actions. During inference, both the video teacher and inverse dynamics model are discarded, leaving only the WAM student for efficient deployment. Simulation and real-world experiments demonstrate that Vid2WAM improves novel-task generalization and data efficiency under limited expert demonstrations while preserving low-latency inference.
Chenhao Qiu, Ruixiang Wang, Runyi Zhao +7
Fudan University · The Chinese University of Hong Kong, Shenzhen · Shanghai Innovation Institute
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.
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