- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper proposes a DualFuture approach for action chunk correction. Without fine-tuning the model weights, the proposed method corrects the predicted actions from a pretrained VLA model. The approach contains three stages. An even-trigger first detects the critical events from an imminent gripper transition. Then, the proposed DualFuture applies two types of future prediction, including a task-consistent future and the action-induced future from the predicted action chunk. Once these two types of future are obtained, the residual derived from these two futures is used to correct the original action chunk. The experiments are conducted on RoboCasa365 and the real-world Franka robot.
Critical strengths
- Action correction without fine-tuning the VLAs are practical for real-world deployment. It preserves the knowledge learned from large-scale pretraining while providing an efficient way to refine the proposed action chunk before execution.
- The proposed method is reasonable and technically sound. It leverages the latent future prediction as signals to guide the refinement of the action chunk.
- The overall paper is well structured.
Critical weaknesses
- The proposed DualFuture adds additional modules upon a frozen VLA model, which could increase inference costs. As real-time execution is critical for robot manipulation tasks, it would be more comprehensive that this paper could also provide the experiment to analysis and discussion about the trade-off between performance and the induced additional inference costs.
- For the future prediction, this paper applies V-JEPA-2 as the prediction target. The motivation behind the use of feature-space future prediction instead of in the pixel space are not clearly highlighted. In addition, for feature-level prediction, previous works also adopt other foundation models like DINO features, etc as the prediction target. Providing the clarification or comparisons among different future prediction targets would be beneficial.
- The Figure 2 is not easy to read due to the orientation of the text in the figure.
Questions for authors
- Is the proposed method model-agnostic? The current method is based on pi0.5. Whether the proposed approach can be applied to different foundation VLA models like GR00T N1.5 or RLDX?
3 comments
Response to Reviewer E3b7 - Part 2
[Question 1] Transfer Across VLA Backbones
DualFuture is designed to work with different VLA backbones. It takes the action chunk proposed by a frozen policy and predicts a residual correction. To test whether a corrector learned on one backbone remains effective on another, we evaluated DualFuture with GR00T N1.5 on RoboCasa365 Composite Unseen under the same protocol used for π0.5.
Applying the π0.5-trained DualFuture to GR00T N1.5 without retraining raised task success from 44.0% to 48.875%, a gain of 4.875 percentage points. When we trained DualFuture specifically for GR00T N1.5 while keeping the backbone frozen, success reached 49.38%, a gain of 5.38 points. This is close to the 5.5-point gain observed on π0.5.
Direct transfer thus retained approximately 91% of the gain achieved with backbone-specific training. These results demonstrate transfer from π0.5 to GR00T N1.5.
Thanks to the authors for the response. My concerns have been addressed. I will retain my positive score.
Response to Reviewer E3b7 - Part 1
[Weakness 1] Inference Costs
On RoboCasa365 Composite Unseen, DualFuture improves the success rate of the frozen π0.5 policy from 45.4% to 50.9%, a gain of 5.5 percentage points. On an RTX 5090, DualFuture evaluates its trigger once every 16 observation steps. Each evaluation takes 10.27 ms, equivalent to an amortized monitoring cost of 0.64 ms per observation step. When the trigger fires, future prediction and residual correction take a further 26.69 ms, bringing the total additional latency for that cycle to 36.96 ms. The average intervention rate was 20.96%.
The table compares DualFuture with two baselines under the same trigger schedule, π0.5 backbone, training data, and evaluation protocol. All latency figures represent overhead beyond ordinary base-policy planning. A single π0.5 planning call takes 77.88 ms on the same hardware.
All three correction methods incur the same monitoring cost. Once the trigger fires, Triggered Replan makes an additional π0.5 planning call, Sample-and-Rerank evaluates four candidate action chunks, and DualFuture directly corrects the existing chunk. Among these methods, DualFuture achieves the highest task success rate and incurs the lowest additional latency on triggered cycles.
We also benchmarked the open-source VLA-Corrector on the same RTX 5090. It achieves 46.25% task success and runs detection at every observation step, taking 38.81 ms per step, compared with DualFuture’s amortized 0.64 ms per step. After triggering, VLA-Corrector’s gradient-guided recovery takes 145.75 ms, compared with 26.69 ms for DualFuture’s future prediction and residual correction.
For context, FPC-VLA reports a correction latency of 1.766 s per keyframe using Qwen2.5-VL-7B as its supervisor. VLA-in-the-Loop reports 21 s per correction using a 7B discriminator and a 14B video generation model. These figures come from the respective papers.
[Weakness 2] Choice of Latent Representation
DualFuture predicts both the expected outcome of the current action chunk and the task-desired state in a shared feature space, enabling the corrector to directly correct actions based on the difference between the two, without reconstructing future images. We prioritized V-JEPA 2 because its features have been validated in real-robot planning: V-JEPA 2-AC predicts the consequences of candidate actions in this feature space and uses them to perform pick-and-place tasks on a Franka robot arm. This aligns directly with DualFuture's need to correct actions based on predicted consequences.
We further replace the prediction target with DINOv3 ViT-L/16 features while keeping the same training and evaluation setup. Surprisingly, we observe that although DINOv3 ViT-L/16 has only about 30% of the parameters of V-JEPA 2 ViT-g/16 (300M vs. ~1,000M), DualFuture with DINOv3 features achieves a higher success rate on RoboCasa 365 Composite Unseen (51.75% vs. 50.9%). Both feature targets improve on the frozen π0.5 baseline (45.4%), with gains of 6.35 percentage points for DINOv3 and 5.5 points for V-JEPA 2. These results show that pixel space DINOv3 features can also serve as an effective prediction target for action correction in DualFuture.
[Weakness 3] Size of Figure 2 Text
Thank you for pointing this out. We will follow your advices to revise Figure 2 so that all labels can be read horizontally without rotating the page .