Video Prediction Policy 2: A 14B Model Beats π0.5 by 18.5 Points on Zero-Shot ALOHA (2610.10270)

Video Prediction Policy 2 (VPP2), arXiv 2610.10270, reports 58.5% average success on 10 real-world zero-shot ALOHA task categories, against 40.0% for π0.5 and 20.5% for Fast-WAM. That is an 18.5 point lead over the strongest baseline. The paper, by Yanjiang Guo and 17 co-authors, was submitted on 7 October 2026 (v2 on 8 October), and all numbers below are the authors' own results.

What is Video Prediction Policy 2?

VPP2 is a world action model (WAM): a robot policy that reuses the priors of a video-generation model to predict what should happen next, then turns that prediction into actions. It builds on Wan2.1-I2V-14B, adds a 0.9B-parameter action expert, and targets "strong zero-shot generalization in both video prediction and action generation," according to the abstract.

It follows the original Video Prediction Policy, which fine-tuned a video diffusion model and learned an inverse dynamics model on its internal representations.

What problem does VPP2 try to solve?

The authors say existing WAMs "frequently produce incorrect motion predictions in open-ended environments, leading to erroneous actions." They give two causes: base video models are not optimized for manipulation, and naively adding action components to a video model can substantially degrade its generalization.

So VPP2 attacks both: better video pretraining first, then an action module attached in a way meant to preserve the video model's behavior.

How does VPP2 work?

VPP2 trains in three stages, per the paper's HTML version.

1. Event-level continued pretraining. Each training example is a full manipulation subtask with frames sampled across the whole event. The model predicts 49 frames at 416x240 for 30,000 steps with batch size 1,024, learning rate 1e-5 and a flow-matching loss. Clips run 1 to 12 seconds. Boundaries come from hand or end-effector velocity minima and gripper open/close moments, then a vision-language model refines them and writes captions.
2. Chunk-level post-training and distillation. The model predicts a fixed chunk (2 seconds for human hands, 8 seconds for robots), then is distilled by consistency distillation into a single-step planner. The authors report about 0.12 seconds for one forward pass.
3. Action module. A 0.9B diffusion transformer sits beside the video model in a mixture-of-transformers, conditioned on the video model's KV cache, and learns what the authors call an implicit inverse dynamics model. It uses 5 denoising steps for a 2-second action chunk, about 0.1 seconds, for roughly 0.22 seconds in total.

The training mix covers more than 20 robot embodiments, and the authors say egocentric bimanual data (ALOHA-style, humanoid, human egocentric) makes up more than 80% of it. Sources include OXE, DROID, AgiBot World-Beta, EgoDex, Ego4D and some self-collected ALOHA data.

What are the results?

The authors report three groups of results: video prediction, real-world ALOHA, and simulation benchmarks after benchmark-specific post-training.

Real-world zero-shot ALOHA (10 task categories, average success):

The paper says VPP2 performs best in 9 of the 10 categories. Per-category numbers appear only in a figure.

Video prediction instruction-following (50 image-instruction pairs per domain, GPT-based evaluator):

The abstract's claim of an 11.0 point lead over Cosmos3-64B matches the gap between the two-column averages (0.85 against 0.74).

Simulation benchmarks (after benchmark-specific post-training):

The RoboDojo baselines come from the official leaderboard as of September 2026, and VPP2's RoboDojo numbers use full-task instructions without subtask planning. With a subtask planner on 5 task groups (50 trials per task), the authors report 57.6% average success against 27.6% for the baseline.

Does the evidence hold up?

The direction is plausible, but several details limit how far the headline travels.

• "Zero-shot" means zero task fine-tuning, not a new robot. The paper says the policy ran on "a robot embodiment seen during training, without task-specific fine-tuning." Self-collected ALOHA data is part of the training mix.
• The baselines are the authors' reruns. π0.5 and Fast-WAM were fine-tuned on all ALOHA datasets in VPP2's training data. That is a fair design, but it is not the same as comparing against published checkpoints, and it was done by the team that built VPP2.
• The sample is small and mostly hidden. The comparison averages 10 task categories. The main text gives per-category scores only in a figure, and a 40.0% versus 58.5% average from a handful of categories can move by several points with one or two categories changing. Confidence intervals are not reported in the main text.
• The video metric uses an automated judge. The 50 pairs per domain are scored by a GPT-based evaluator. Human pairwise win rates are also shown, but only as a figure.
• Benchmark selection. LIBERO-PRO argues that models scoring over 90% on standard LIBERO can fall sharply under perturbations. VPP2's own table fits that pattern: π0.5 scores 96.9 on LIBERO-ID but 11.0 on LIBERO-Pro. Comparisons on perturbed suites are informative, but VPP2 was post-trained for each benchmark, so the margin reflects the full pipeline, not only the pretrained model.
• No independent replication yet. The paper was submitted six days before this article's date.

How does VPP2 compare with related work?

The contrast with Fast-WAM is the interesting one. Fast-WAM asks whether a WAM needs test-time future imagination at all. VPP2 keeps a visual plan at inference but makes it a single step, so it argues that the quality of the prediction matters, not whether one exists. In VPP2's ALOHA table, Fast-WAM scores 20.5% against 58.5%, but the two systems differ in more than that one design choice, so the gap does not isolate the cause.

What does a working researcher take from this?

• Event-level prediction targets matter in the video evaluation. The ablation "VPP2-Fixed-Step", which predicts a fixed-duration chunk rather than a complete event, scores 0.48 human and 0.42 robot against 0.80 and 0.90 for full VPP2. That is a video-prediction result, not a real-world robot result.
• Check latency on your hardware. The authors report about 0.22 seconds per 2-second action chunk, and the main text does not name the hardware.
• Check the embodiment. If your robot is not in the training mix, the ALOHA result does not tell you what to expect.
• Read the figure, not the average. Per-category results decide whether the lead is broad or driven by a few tasks.

What's still unclear?

• How many trials were run per ALOHA task, and how large the per-category variance is.
• How much of the ALOHA gain comes from the video pretraining, the distillation, or the MoT action module, since the main text does not provide a full ablation of the real-world result.
• Whether the lead holds on an embodiment outside the training mix.
• Whether the released code and checkpoints reproduce the numbers. The authors say they are on the project website at robert-gyj.github.io/video-prediction-policy-2.
• Whether the paper's limitations are discussed. The conclusion in the main text summarizes results and points to preserving "generalizable video predictions" during action learning, and does not list limitations.

We read: the full main text of the paper (arXiv HTML v2). The appendices were not part of this read, so claims here rest on the main text, tables and abstract.