Best Physical AI Papers of October 2026: World Action Models Take Over

Three new papers push world action models (WAMs) in different directions this month. Long-WAM from NVIDIA and partners reports that 19.2 seconds of video history lifts RoboCasa GR-1 success from 63.3% to 78.7%. ProWAM reports 85.8% on LIBERO-Plus, and VPP2 reports an 18.5 point lead over the strongest baseline on zero-shot ALOHA tasks. All of these are the authors' own numbers, on different benchmarks, with no independent replication yet.

What is a world action model?

A world action model is a robot policy built on a video model that predicts future world states and actions together. DreamZero (February 2026) states the premise: unlike VLAs, WAMs learn physical dynamics by predicting future world states and actions, using video as a dense representation of how the world evolves. The authors report over 2x better generalization to new tasks and environments than state-of-the-art VLAs in real-robot experiments, with a 14B model running closed-loop at 7Hz.

That is why October's papers cluster around one question: how do you keep the benefit of video prediction without paying for full video generation at every step? Each paper below answers it differently.

What does Long-WAM claim?

Long-WAM (arXiv 2610.10528, submitted 7 October 2026) claims that longer visual history only helps if the video foundation model was pretrained autoregressively. On RoboCasa GR-1, raising context from 0.0 to 19.2 seconds lifts success from 63.3% to 78.7%. A bidirectionally pretrained initialization shows no net gain, per the abstract.

Other lab-reported results from the paper:

• LIBERO-Long: 2.4 seconds of history improves success from 94.5% to 99.5%.
• RoboTwin 2.0: 94.4% average success. DOMINO: 34.9%.
• Latency: 107.4 ms per action chunk on an RTX 5090, including future-video latent prediction. It also runs on DGX Spark and Jetson AGX Thor.
• Real robot (Unitree G1): 95% on dynamic cup stacking, where π0.5 and Fast-WAM succeed in none of 20 trials.

The 20-trial count matters. With 20 trials, one run is 5 points, so the gap to a baseline at zero is large but the 95% itself is uncertain. Our earlier Long-WAM write-up covers the memory mechanism in more depth.

What does ProWAM change?

ProWAM (arXiv 2610.02508, submitted 1 October 2026, from Shanghai Jiao Tong University, SII, Meta and Imperial College London) predicts actions plus an ordered sequence of sparse visual sub-goals instead of a dense video rollout. One video-backbone forward pass caches the sub-goal features, so replanning needs only lightweight action denoising.

The authors report state-of-the-art results on LIBERO-Plus (85.8%) and randomized RoboTwin (75.7%), with relative gains of up to 35.9% over the strongest baseline. Zero-shot real-world success is 70.0% against 55.0% for the strongest baseline in novel scenes. The HTML tables show the LIBERO-Plus spread:

Fast-WAM at 51.5 sits below π0.5, a reminder that "WAM" is a design family, not a performance tier. On RoboCasa365, ProWAM scores 48.1% overall and 18.2% on the Composite-Unseen split, and the abstract says it ranks 4th overall. We covered the record claim and its gaps in ProWAM Reports a Record on LIBERO-Plus.

What does VPP2 add?

Video Prediction Policy 2 (arXiv 2610.10270, submitted 7 October 2026, revised to v3 on 13 October) argues that existing WAMs often predict wrong motion in open-ended settings because base video models are not tuned for manipulation. The team from Robotera, Tsinghua and partners continues pretraining on a curated manipulation video set with event-level captions, then distills the video model into a single-step visual planner. A 0.9B diffusion transformer action module learns implicit inverse dynamics.

Reported results from the paper:

Inference takes about 0.22 seconds per action chunk, split roughly between video (0.12 s) and the action expert (0.1 s). See our VPP2 explainer for the architecture.

Where does H-JEPA fit?

H-JEPA (arXiv 2610.06805, submitted 5 October 2026, with Yann LeCun among the authors) is not a WAM, but it competes for the same planning role. It trains a hierarchy of action-conditioned JEPAs where each level predicts farther ahead in its own latent space, and plans top-down. On Visual AntMaze, a three-level hierarchy raises success from 18% to 73% with less planner compute, according to the authors. The real-robot evidence is weaker: on DROID videos the paper reports improved offline planning fidelity, not task success.

The contrast with the WAM papers is the point. H-JEPA plans without generating pixels, while WAMs bet that predicted video is a useful intermediate. We compared the two views in LeCun vs Video Generation.

Does the evidence hold up?

Partly. Four weaknesses apply across the batch.

1. Different benchmarks. Long-WAM reports RoboCasa GR-1, ProWAM reports LIBERO-Plus, VPP2 reports RoboDojo and ALOHA. You cannot rank the papers by their headline numbers.
2. Baseline choice. Each paper picks baselines and, in VPP2's case, fine-tunes π0.5 and Fast-WAM on the same ALOHA data. VPP2 also notes the ALOHA embodiment was seen in training, so "zero-shot" refers to tasks, not the robot.
3. Small real-robot samples. Long-WAM's decisive comparison uses 20 trials for the baselines, and its abstract does not state Long-WAM's own trial count. VPP2 reports no trial count per ALOHA category.
4. Mixed standing. ProWAM leads LIBERO-Plus, but its abstract ranks it only 4th overall on RoboCasa365.

A broader check comes from the study: Cosmos-Policy reaches 82.2% on LIBERO-Plus and LingBot-VA 74.2% on RoboTwin 2.0-Plus, yet the authors say VLAs such as π0.5 can match WAM performance on certain tasks, though they typically need extensive, varied training.

What does a working researcher take from this?

Three practical points follow from the reported results.

• Pretraining style gates the value of context. Long-WAM's central claim is that longer histories pay off far more when the video foundation is pretrained autoregressively. If you are adding memory to a policy, test that before scaling context.
• Dense video generation is becoming optional. ProWAM caches sparse sub-goals, and VPP2 distills to a single-step planner. Both aim to cut inference cost while keeping the prediction prior. Our earlier note on first-denoising-step WAMs points the same way.
• Check the code before committing. Long-WAM and VPP2 list code repositories, and ProWAM links a project page. Check each for trained weights. ProWAM's paper is licensed CC BY-NC-ND 4.0, which governs the paper text, not necessarily the code.

What's still unclear?

Several questions remain open as of 15 October 2026.

• No paper has been replicated by an outside group. Every number above is lab-reported.
• It is unknown whether Long-WAM's memory gain transfers to tasks where history is not the bottleneck. LIBERO-Long peaked at 2.4 seconds while GR-1 peaked at 19.2 seconds.
• ProWAM's real-world result is a gain of 15 points over the strongest baseline, but the abstract does not say how many trials back it.
• Whether single-step planners such as VPP2's keep their generalization on embodiments unseen in training is not shown, since ALOHA was seen in training.
• The H-JEPA versus WAM question (plan in latents, or in predicted pixels) has no head-to-head benchmark yet.

Read this month's papers as a clear direction, cheaper prediction with better priors, rather than a settled ranking.