acceptodds
Withdrawn from ICLR 2027

SWAM: Self-Supervised World Action Modeling via Asymmetric Denoising

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We revisit denoising in world action models from a self-supervised representation learning perspective, shifting emphasis away from video generation. Rather than relying on costly, error-prone future-video rollouts, robot policies need robust visual features that support action prediction from current observations. We therefore introduce the Self-Supervised World Action Model (), which trains action-relevant robust visual features through asymmetric denoising. uses Denoising Autoencoding (DAE) on current observations with a noise schedule less aggressive than that of video generation, providing auxiliary self-supervision for the visual backbone. For future observations, Masked Autoencoding (MAE) motivates extreme-noise denoising that limits access to the target and enables single-pass visual inference. achieves success rates of 99.1% on LIBERO and 83.4% on LIBERO-Plus. On LIBERO-Plus, it exceeds Fast-WAM and Fast-WAM-Joint by 31.9 and 15.0 percentage points, respectively, with comparable inference-path latency to Fast-WAM and over 50% lower latency than Fast-WAM-Joint. These results support asymmetric denoising as an effective self-supervised objective for robot policy learning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Action Model, robot policy, self-supervised representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4cZ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SWAM, a world-action model that uses asymmetric visual denoising to learn representations for robot control. An attention-isolated auxiliary branch denoises corrupted current observations using shared visual weights, while the future branch is trained with noise concentrated near the pure-noise endpoint. At deployment, the visual backbone processes the current observation and noisy future slots once, and the resulting features are cached for iterative action denoising. Evaluations on LIBERO, LIBERO-Plus, RoboTwin 2.0, and four real-robot tasks show improvements over Fast-WAM, with inference latency close to Fast-WAM and substantially lower than Fast-WAM-Joint.

Critical strengths

1.Substantial empirical improvements on a practically relevant robustness problem. SWAM achieves 83.40% on LIBERO-Plus, compared with 51.50% for Fast-WAM and 68.39% for Fast-WAM-Joint, while maintaining strong standard-LIBERO performance. The real-robot experiments also show an improvement from 59.0% to 76.5% over Fast-WAM across four tasks. These results provide useful evidence that the proposed training recipe improves performance beyond the original training distribution.

2.A coherent design supported by relevant component analyses. Distinguishing available current observations from unavailable future observations is a reasonable motivation for different denoising conditions. The auxiliary current branch updates shared visual weights without adding deployment-time computation. The reported ablations examine current-observation supervision, future-noise sampling, and access to future-position features. In particular, the improvement from 77.18% to 83.40% when adding current-observation denoising supports the usefulness of this auxiliary objective.

Critical weaknesses

1.The methodological contribution appears incremental relative to closely related work, and the most relevant comparisons are missing. Single-pass construction of future-conditioned features and their reuse during action denoising are already central to ForeWAM, Faster-WAM, and RIFT. Likewise, treating denoising as representation learning has established precedents in Deconstructing Denoising Diffusion Models for Self-Supervised Learning and DiffMAE. SWAM acknowledges these connections, but its main experiments do not compare against the closest single-pass WAM methods. The distinctive contribution therefore appears to lie primarily in the particular combination of auxiliary current denoising, future-noise sampling, and loss weighting. Stronger evidence is needed to establish the value of this combination beyond existing future-feature interfaces.

2.The experiments do not fully isolate the proposed asymmetry or establish the stability and generality of its benefits. The ablations are useful, but multiple factors change relative to Fast-WAM: access to future features, future-loss weighting, future-noise sampling, and current-observation supervision. Notably, Figure 4 reports a 5.48-point gain from changing future-loss weighting, whereas changing the sampling distribution from Uniform to Beta(7,1) adds 1.57 points under uniform weighting. This suggests that the contribution of the proposed noise asymmetry should be distinguished more carefully from other training changes. Matched comparisons between symmetric and asymmetric noise assignments, with auxiliary supervision and loss scaling controlled, would strengthen the interpretation. Moreover, the main results lack variability estimates across independent training runs. The robustness claim should also be qualified: RoboTwin randomized success improves over Fast-WAM but remains only 15.2%, below several other reported methods.

Questions for authors

1.How does SWAM compare with ForeWAM, Faster-WAM, or RIFT under matched backbone, training-data, optimization-budget, and evaluation conditions? What benefit can be attributed specifically to the asymmetric denoising recipe beyond the shared single-pass future-conditioning interface?

2.Can you isolate the effects of noise sampling and loss weighting while keeping the architecture and auxiliary objectives fixed? In particular, how do symmetric noise assignments compare with the proposed asymmetric assignment under controlled loss scaling?

3.In the future-feature readout diagnostic reporting 77.18% versus 62.14%, was the model without future access retrained under the corresponding attention restriction, or was access removed only at evaluation? This distinction affects whether the result establishes a training benefit or sensitivity to an inference-time intervention.

4.How reproducible are the main improvements across independent training seeds? Do the contributions of current denoising and future-noise scheduling remain consistent on RoboTwin or the real-robot tasks, beyond the LIBERO-Plus ablations?

Reviewer_BfEWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SWAM, a world action model that uses asymmetric denoising to learn visual representations for robot control. The method uses two complementary training objectives: an auxiliary denoising task on current observations and a heavily corrupted future prediction task, rathher than focusing on generating future videos. This allows the policy to use both current and future-position features while requiring only one visual forward pass during inference. Experiments on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world manipulation tasks show substantial improvements over Fast-WAM, particularly under distribution shifts.

Critical strengths

First the paper has a clear and well-motivated idea. I really like the observation that robot policies do not necessarily need accurate future video generation, but rather useful representations for action prediction. The distinction between current observations and unavailable future observations is simple but meaningful. The proposed asymmetric denoising strategy follows naturally from this motivation.

Second the practical balance between performance and inference cost is meritorious. SWAM preserves future-position information without requiring iterative video generation at inference. Although it is slightly slower than Fast-WAM, the latency remains much lower than Fast-WAM-Joint. This seems like a reasonable trade-off given the substantial performance improvements.

Critical weaknesses

Well, fisrts the methodological novelty is somewhat limited. The main components build on existing denoising autoencoders, masked autoencoding, and the Fast-WAM architecture. The contribution mainly comes from adapting the noise schedules and combining the two objectives for policy learning. While the results are strong, I would have liked a clearer discussion of what distinguishes this approach from other forms of auxiliary visual supervision.

Also robustness improvements are not equally strong across all benchmarks. SWAM significantly improves over Fast-WAM on RoboTwin Randomized, but its success rate remains relatively low at 15.2%, compared with 25.8% for X-WAM and 30.4% for ABot-M0. These methods use different pretraining settings, so the comparison is not entirely direct. Still, it would be useful to better understand why the strong improvements on LIBERO-Plus do not translate into similarly competitive results on RoboTwin Randomized.

The contribution of different design choices could also be separated more clearly. The paper provides useful ablations on noise schedules and the auxiliary current-observation loss. However, the improvement over Fast-WAM also involves retaining future-position tokens at inference. I would like to see a more controlled comparison that separates the benefits of the asymmetric denoising objectives from the benefits of adding future-position features.

Questions for authors
  1. Could the authors provide a more controlled comparison between the asymmetric denoising objectives and the use of future-position features? In particular, how much of the improvement comes from the proposed training objectives when the architecture and inference procedure are kept identical?

  2. What might explain the relatively large gap between LIBERO-Plus and RoboTwin Randomized performance? Are there particular types of perturbations or manipulation tasks where the learned representations still struggle?

  3. Have the authors evaluated the stability of the results across different random seeds? Reporting variance for the main benchmarks would help establish how consistent the improvements are.

Reviewer_FqLvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SWAM, building on Fast-WAM by adding an auxiliary self-supervised objective to improve manipulation performance on both simulation and real-world tasks. The paper evaluates its method on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world manipulations, outperforming Fast-WAM and previous approaches.

Critical strengths

S1: The paper improves Fast-WAM using a self-supervised objective.

Critical weaknesses

W1: The impact of the paper seems limited to me. The method is only evaluated on manipulation tasks and is also limited to offline world modeling with datasets of episode trajectories.

W2: The paper does not show the impact of removing the current observation loss on LIBERO, RoboTwin 2.0, and real-world tasks. It only shows the impact of the contribution on LIBERO-Plus.

W3: The obtained results on RoboTwin 2.0 appear significantly lower than the results obtained in the Fast-WAM paper (Table 1). Fast-WAM obtains an average performance of 37.8 instead of 91.8 in the official paper.

Questions for authors

Q1: What is the impact of removing the current observation loss on LIBERO, RoboTwin 2.0, and real-world task performance?

Q2: What is the reason for the inconsistency of RoboTwin 2.0 results between the authors' paper and the official Fast-WAM paper?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.