- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates whether World Action Models (WAMs) truly learn causal dynamics through future prediction. The authors observe that existing WAMs are insensitive to temporal shuffling of predicted frames and changes in action conditions. To address this, they propose a two-stage training method using multimodal failure trajectories for dynamics learning and expert trajectories for policy learning. Experiments show improved future prediction, particularly on failure trajectories.
Critical strengths
-
The paper studies an interesting question of whether WAMs truly capture causal dynamics. Experiments identified that existing WAMs are insensitive to temporal shuffling of predicted frames and changes in action conditions. The motivation is well-supported.
-
The experiments show that failure-pretrained models can better predict future trajectories under unsuccessful actions. Under controlled condition, main results show the superiority of the failure-pretrained method over unimodal training.
Critical weaknesses
-
The action-conditioned model does not show an advantage over the model without action conditioning. Could this be due to insufficient training data? I think this possibility should be ruled out.
-
The motivation behind the proposed method is reasonable, but the experiments do not directly validate it. It would be helpful to evaluate the failure-pretrained model under the settings of Tables 1 and 2 to further assess its effectiveness in learning causal dynamics. Could the authors provide the results?
Questions for authors
The necessity of two-stage training is not sufficiently justified. Could the authors provide ablation studies comparing the proposed method with alternative training strategies?