- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents GE-Act 2.0, a world action model that is trained from scratch instead of being initialized from a pretrained video diffusion model. Specifically, they train a new autoencoder called CoAE, a single-step visual planner (SVP; analogous to a video diffusion model), and an IDM (analogous to a policy/action predictor). For CoAE, they regularize the latent space to be aligned with pretrained visual representations; for SVP, they adopt MeanFlow to enable one-step generation; and for IDM, they predict actions from the current and predicted future frames. Thanks to the compact latents and one-step generation, they achieve real-time inference. They first independently train the IDM using GT video-action pairs and the SVP using GT videos, and then co-train both models by feeding generated rollouts to the IDM and pairing them with GT actions. However, generated rollouts sometimes do not match the GT actions, so the authors propose KASO, which selects the rollout that best matches the GT actions among multiple rollouts from the SVP. The selection is based on the IDM’s velocity differences (i.e., energy) between the GT recorded future and each generated candidate. By doing so, they mitigate this issue and show zero-shot OOD scaling and fine-grained instruction-following capabilities, as well as competitive benchmark performance.
Critical strengths
- The paper is generally well-written and easy to follow.
- Instead of initializing from the pretrained video diffusion models, the paper tries to do more comphrehensive study by training every model from scratch.
Critical weaknesses
- While I know the terms “WAM” and “VAM” are used interchangeably these days, I think the term “WAM” may be misleading in this paper. Specifically, many WAMs jointly predict future video frames and corresponding actions during both training and inference, in which case the validity gap described in the paper may not arise. I think this issue mainly stems from the fact that the paper trains a video model and a decoupled action predictor. In this respect, the paper should clarify that the problem is specific to decoupled architectures, and ideally discuss or compare with joint-denoising approaches.
- I think 10 trials per target is not enough, especially for the KASO ablation in Table 6, and more trials are needed to reduce the noise. For example, with 10 trials, a 40% success rate has a 95% confidence interval of roughly 17–69%, which makes the differences between methods hard to interpret.
- For the scaling experiment in Table 1, I think the number of training samples seen should be identical across models, rather than training each model for one epoch. With the current setup, the number of iterations for the 0.3k h model is 100× smaller than for the 30k h model, so the effects of data scale and training compute are confounded.
Questions for authors
- Could the authors provide real-world experiments with more trials to reduce the noise?
5 comments
An Additional Zero-shot Comparison
Dear Reviewer zP4W,
In addition to the experiments in our first reply, we added an experiment that directly compares the zero-shot ability of GE-Act 2.0 with a VLA trained under matched conditions. We trained StarVLA [4] on the same 30,000 hours of co-training data as GE-Act 2.0, with the same Qwen3.5-2B vision-language backbone. Both models were evaluated zero-shot, without task-specific fine-tuning, on the same G1-OP robot in the same unseen environments. To cover every skill while keeping real-robot evaluation tractable during the rebuttal period, we evaluated both models on a subset of 47 of our 100 tasks, taking about two tasks from each skill category (all tasks where a category has two or fewer), so that all 20 skill groups are represented. This subset is not easier for GE-Act 2.0: its success on these 47 tasks (36.0%) is below its average over all 100 tasks (44.1%). Skill groups follow Table 1.
| Skill group (tasks) | StarVLA | GE-Act 2.0 | Skill group (tasks) | StarVLA | GE-Act 2.0 |
|---|---|---|---|---|---|
| Pick (5) | 54.0 | 76.0 | Close (1) | 70.0 | 90.0 |
| Place (2) | 80.0 | 100.0 | Separate (2) | 15.0 | 25.0 |
| Sweep + Wipe (2) | 80.0 | 40.0 | Press (2) | 15.0 | 30.0 |
| Stack (2) | 20.0 | 20.0 | Pull (3) | 0.0 | 13.3 |
| Pass (2) | 80.0 | 85.0 | Hang (2) | 35.0 | 0.0 |
| Fold (2) | 20.0 | 25.0 | Tear (2) | 0.0 | 5.0 |
| Straighten (2) | 50.0 | 35.0 | Flip (3) | 0.0 | 30.0 |
| Pour (2) | 0.0 | 45.0 | Zip (2) | 0.0 | 0.0 |
| Push (4) | 27.5 | 30.0 | Shake/Stir (4) | 0.0 | 12.5 |
| Insert (3) | 43.3 | 33.3 | All 47 tasks | 29.1 | 36.0 |
Over all 47 tasks, GE-Act 2.0 reaches 36.0% zero-shot success (95% CI 32 to 40) against 29.1% for StarVLA (25 to 33). The 6.8-point difference is significant under a Cochran-Mantel-Haenszel test stratified by skill group (p = 0.008). GE-Act 2.0 is higher on 13 compared groups and lower on 4. With 470 trials per model, this comparison also adds to the larger-scale real-robot evidence you asked for.
Again, thank you very much for your suggestions! We would deeply appreciate it if you could let us know whether these additional results address your concerns.
References
[4] StarVLA Community. StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing. arXiv:2604.05014, 2026.
Thanks for the rebuttal. I still have a concern in data scale and training compute: I think the authors should provide 100 passes over the 0.3k-hour data for better comparison. I think in recent imitation-earning-based polices in robotics behave quite differently from other domains like LLM that you mentioned, e.g. https://seohong.me/blog/behavioral-cloning-mystery/ . So I think the data scaling claim may be too strong.
Dear Reviewer zP4W,
We sincerely thank you for the careful and constructive review! Your comments pointed us to the places where the paper needed more evidence and clearer wording. Below, we address your concerns point by point.
W1. The term "WAM" and joint-denoising approaches. Thank you for this precise observation on terminology. You are right that many WAMs jointly predict future video frames and actions, and that the validity gap may not arise there. Following your suggestion, we added a clarification in Section 2.3 that the problem is specific to decoupled architectures.
On joint-denoising approaches: our related work (Appendix A) separates decoupled two-stage designs from joint designs. Joint-denoising models avoid the gap because video and action are denoised against the same recorded sample. In a decoupled architecture, the gap appears once the video model and the action predictor are co-trained: the action predictor receives a generated future but is supervised with the recorded action. This is the setting we study.
W2 and Q1. More trials to reduce the noise. Thank you for this helpful suggestion. We agree that 10 trials per target is not enough, especially for the KASO ablation in Table 6. We added real-world experiments with more trials to reduce the noise: a second full batch of the whole ablation under the same protocol. Together with the original trials, this gives 50 single-object trials and 20 trials per target in the four-object scene, which is 130 pick trials per method.
With the added trials, the differences between methods are now clear to interpret. The 95% interval for KASO narrows from the 17 to 69 you computed for 10 trials to 33 to 49. KASO is 19.3 points above E2E+PT and 16.2 points above E2E in pooled pick success, and both differences are significant (Cochran-Mantel-Haenszel test stratified by protocol, p = 0.0008 and p = 0.006).
W3. Data scale and training compute in Table 1. Thank you for looking so closely at our scaling protocol. Separating the effect of data from the effect of compute is exactly the right question to ask of a scaling study, and we are glad to explain our choice. On whether the number of training samples seen should be identical across models: we follow the common convention in data-scaling studies, where training length is set by the data and a larger dataset naturally receives more training steps. Large language models are trained in about a single pass over their data, so more data means more steps [1]. In robot learning, data-scaling studies likewise give larger datasets more training steps so that every size converges [2].
We train each model for one epoch because our internal evaluation showed that, at around one epoch, the model has converged and is well trained without overfitting. Making the number of samples seen identical would mean about 100 passes over the 0.3k-hour data. Repeating data that many times is known to bring no further gain and introduces overfitting at the current data scale [3]. We hope this explanation clarifies our choice of scaling setup.
Again, thank you very much for your suggestions for improvements! We would deeply appreciate it if you could let us know whether your concerns have been addressed properly.
References
[1] Hoffmann et al. Training Compute-Optimal Large Language Models. NeurIPS 2022.
[2] Lin et al. Data Scaling Laws in Imitation Learning for Robotic Manipulation. ICLR 2025.
[3] Muennighoff et al. Scaling Data-Constrained Language Models. NeurIPS 2023.