The work struggles to clarify what the reader learns from it. This isn't necessarily bad, since it is a systems contribution presenting an overall system that achieves state-of-the-art performance, so it shouldn't necessarily be held against the authors. Even for systems-level work, the authors need to clarify which design choices were obvious, what was challenging, and what new insight led to this performance. I.e., the work needs to tell its readers why it is interesting from a research perspective.
As it stands, the work is unreproducible, which is a deal-breaker. I double-checked this with an LLM and will list its findings below verbatim. I tried to filter the incorrect ones out and double check as good as I could. Even after checking it still seems that numerous points are missing for the work to be reproducible.
This is less of a critical weakness but more of a point for potential improvement. While the writing is very clear, it is largely only understandable to researchers who already work in robot learning. Even basic knowledge in either Computer Vision or Robotics isn't enough to understand this paper, and it is completely out of reach for the average ICLR attendee, where the common knowledge denominator assumption is a grad-level deep learning/ML course.
LLM output on missing information for reproducibility
Data
- Corpus size is missing: clips, hours, or frames per source in Table 1, the total, and the mixture or sampling weights across sources.
- The self-collected ALOHA, dexterous-hand, and egocentric data have no stated amount, tasks, hardware, or cameras.
- The per-dataset filtering strategies (Table 1 caption) and the general-video filter ("similar to LVP") are not specified: criteria, models, and thresholds are all missing.
- For segmentation, the paper doesn't give:
- the velocity and gripper-event thresholds
- how hand velocity is obtained for human videos
- which VLM refines the segments, and its prompt
- For captioning, the paper doesn't give:
- the VLM identity and version
- the full caption template or prompt
- the optical-flow method and threshold for discarding egocentric clips
- The unified action space is missing:
- the action parameterization: absolute vs. delta end-effector pose, rotation representation, gripper/hand encoding
- action dimensionality, normalization, and control frequency
- whether proprioception is an input
- how dexterous-hand, single-arm, human-hand, and non-egocentric (~20%) data are mapped or excluded
Stage 1 (event-level pretraining)
- The optimizer and its settings are missing: LR schedule and warmup, weight decay, gradient clipping, EMA, and precision.
- The flow-time sampling distribution (uniform, logit-normal, or shifted) is not stated.
- The text-dropout rate for CFG is not stated.
- The paper doesn't say which components are trainable, e.g. whether the text encoder, CLIP image encoder, and VAE are frozen.
- There's no description of how events shorter than N=16 frames are handled, or how differing source frame rates are handled.
Stage 2 (chunk post-training and consistency distillation)
- Post-training steps, batch size, LR, and data mixture are not given.
- For consistency distillation, the paper doesn't give:
- the weighting λ(s, s′)
- how F_θ is parameterized in terms of v_θ
- the teacher ODE solver and step count used to compute x̂_{s′}
- the sampling distributions for s and for s′ < 1
- the teacher CFG scale and the EMA decay
- steps, batch size, LR, and student initialization
- The multi-step sampler and CFG scale behind the 30-step results in Fig. 5 are not stated.
Stage 3 (action expert)
- The 0.9B DiT architecture is undescribed: depth, width, heads, and how its layers pair with Wan's blocks in the MoT.
- The action chunk length in steps is missing (only "2 seconds" is given), as are the action tokenization, action loss, and noise schedule.
- It's unclear whether the video loss is kept during Stage 3, and what the loss weights are.
- It's unclear which video latents condition the actions during training: ground truth, noised, or one-step distilled outputs. This matters because inference uses the one-step output.
- The freeze/LoRA schedule is missing:
- how long the backbone stays frozen
- LoRA rank, alpha, and target modules
- whether full unfreezing follows
- total steps, batch size, and LR
- Which datasets enter action training, and in what mixture, is not stated.
- Deployment details are missing: actions executed per replanning step, control rate, and the GPU used for the 0.22 s latency figure.
VLM planner and prompt enhancement
- The paper doesn't say which VLM is used in each experiment (both Qwen3-VL and Seed2.0 are cited).
- The planning and prompt-enhancement prompts are not given, nor how "execution feedback" and subtask completion are determined.
- It's unclear whether prompt enhancement or planning was used in Table 2, the ALOHA experiments, and LIBERO, and whether baselines received the same enhanced prompts.
Evaluation
- Video prediction:
- The 50+50 image–instruction pairs are not released, including the authors' own photos.
- The GPT evaluator's model version, prompt, success criterion, and the frames it was shown are not given.
- The number of samples or seeds per pair is not stated.
- The human-evaluation protocol is missing: number of raters, blinding, and how ties were decided.
- Baseline generation settings (steps, CFG, resolution) are not given.
- It's unclear which VPP2 checkpoint and step count were used, and how VPP2-Fixed-Step was trained.
- ALOHA zero-shot:
- The task list, exact instructions, and scene setups are not given.
- Trials per task are not stated (the 5% granularity suggests 20), nor are the success criteria.
- The paper doesn't explain how the tasks were verified as unseen.
- The fine-tuning recipes for π0.5 and Fast-WAM are missing, including their starting checkpoints.
- LIBERO:
- The fine-tuning recipe is missing: starting checkpoint, steps, LR, LoRA vs. full fine-tuning.
- There's no explanation of how single-arm Franka actions map into the bimanual action space, or which camera views are used.
- Episode counts for LIBERO-ID, LIBERO-Pro, and LIBERO-OOD are not given.
- RoboDojo:
- The fine-tuning recipe is missing.
- For subtask planning, the demonstration segmentation and the VLM used are not described.
- VPP2's generalization score is reported only pooled, while baselines are split into Std./Rand.
- No variance across seeds or confidence intervals is reported for any result.
Inconsistencies worth raising
- Video-model latency is about 0.1 s in the Fig. 4 caption but about 0.12 s in Sec. 3.2.
- Fig. 4 depicts joint video-and-action denoising in Stage 3, while Sec. 3 says the action expert conditions on the video model's KV cache. These imply different training procedures.
Thank you for your thoughtful questions about our evaluation.
W1, Q1. Which task instances, objects, and instructions were used in the ALOHA evaluation in Figure 8?
Figure 8 summarizes zero-shot performance across ten real-world task categories on ALOHA. To construct individual task instances, we used everyday objects from our surroundings and placed them randomly in the workspace. This tests whether a policy can carry out the requested manipulation as the objects and their positions vary. Figure 10 shows a random sample of the test tasks and gives concrete examples of the objects and arrangements behind the category-level results. For each task, we used Qwen3-VL [1] to expand the original instruction into a more detailed prompt, which we shared across policies.
W2, Q2. How does event-level pretraining affect policy performance?
We varied only how the training videos were segmented in Stage 1 of VPP2 pretraining. A fixed-length clip can end before the action described by its instruction is complete. Using a complete subtask event keeps the instruction paired with the trajectory through the end of the action. We kept all subsequent training stages identical and evaluated both policies in simulation and on the real robot.
Simulation. Both variants achieved near-perfect success on LIBERO-ID (98.6% with fixed-length clips and 98.8% with complete events), leaving little room for improvement on this benchmark. On the more challenging LIBERO-Pro and LIBERO-OOD benchmarks, event-level pretraining improved success by 4.9–13.2 percentage points in four of the five categories. The largest gain was on OOD Goal, where success rose from 50.2% to 87.7% (+37.5 points).
Real-world zero-shot evaluation. Given the limited time available during the rebuttal, we evaluated both policies on three real-world task categories (stacking, pick-and-place, and towel folding). Event-level pretraining raised the mean success rate across these three categories from 51.7% to 70.0% (+18.3 percentage points). Stacking and towel folding each improved by 10 points, and pick-and-place showed the largest gain, from 40% to 75%. These results show that the gains observed in simulation also extend to the real-world tasks we tested.
Q3. Did all policies in Figure 8 receive the same instructions and VLM prompt enhancement?
Yes. We used Qwen3-VL to expand each task instruction and supplied the same resulting prompt to all three policies (π0.5, Fast-WAM-14B, and VPP2). We followed this procedure during both training and evaluation. In particular, any task details added during prompt enhancement were available to all three policies at test time, so the comparison in Figure 8 holds the language input constant.
[1] Qwen3-VL Technical Report. arXiv:2511.21631, 2025.
Thank you again for your constructive comments and support. We hope these responses address your concerns, and we would be happy to clarify any additional questions.