World-Action Models Agree on the Bottleneck, Not the Fix

Figures attributed to companies and papers are their published claims. The comparisons and judgements below are this author’s analysis.

On 1 October, four submissions landed on arXiv cs.RO within a day of each other: PEWAM, ViGAR, ProAct and NEEDLEWORK, with Awomo-SimDataEngine earlier that morning. Two days earlier Runway announced Praxis-1, an open-weight world action model. All of them begin from one diagnosis, put bluntly by Runway’s chief technology officer Kamil Sindi to The Robot Report: most robot policies are bottlenecked by robot data, which is scarce and expensive to collect.

The answers diverge. Runway’s fix is to stop depending on demonstrations and take physics from video. PEWAM’s fix is to delete the actor from the training signal and keep only what happened to the objects. ViGAR’s fix is a subgoal planner on a shared world representation. ProAct’s is to change the distribution a flow policy samples from. NEEDLEWORK and Awomo leave the policy alone: one rewrites the dataset offline, the other manufactures the simulator worlds that produce it.

That is a disagreement about conditioning, and about what counts as evidence, not about size. It matters if you have to pick one to build on, because the numbers come in a currency that this week’s most careful evaluation says cannot settle the question.

The label stopped carrying information

PEWAM and ViGAR both use the phrase in their titles and solve different problems. Runway uses it for a policy distilled from a video model. ProAct is a vision-language-action paper arguing that a world model should decide where generation starts, not only how it evolves. Awomo builds training worlds for a world action model someone else trains. Six artefacts, one phrase, no shared account of what the phrase buys. What they do share is the constraint: robot data costs robot hours, and every proposal is a way to extract more from data that already exists, the tension described in the two halves of the robot data bottleneck.

Where each bet puts the generality

Video pretraining (Runway)

Runway’s claim is that the ceiling belongs to video rather than demonstrations: performance improves as video scales, so the limit is how much video a model can learn from, not how many teleoperation episodes exist. Praxis-1 is built on the pretraining behind Solaris and GWM Worlds 2 and will ship with open weights, with public release planned for the coming months. The number Runway offers is that simulating robot policies inside its world model predicts real-world results with 0.95 correlation, which it says beats more expensive 3D reconstruction based techniques. That correlation is Runway’s own measurement. Early partners Noble Machines, Standard Bots and Ultra are running one model across bimanual arms and humanoids with light fine-tuning, which is a programme in progress rather than a reported transfer result. The launch was covered in more detail here.

Effect programs (PEWAM)

PEWAM states the problem as ambiguity inside the data. A demonstration records two things in the same frames, the authors write: what happened to the objects, and how one particular arm made it happen. They condition on the first. A demonstration is compiled into an effect program, which is the 3D keypoint trajectories of the objects that moved, two points marking where each was held, and the scene’s final configuration, with the demonstrator removed. PEWAM, a 71.5 million parameter model, generates effect, robot execution, action and terminal state as four streams with independent flow-matching times. Clamping a program and sampling the execution turns inference into something closer to programming, re-solved closed loop from the live scene.

The results are the most specific in the set. On held-out LIBERO-Goal tasks, one demonstration’s program completes 40 of 90 episodes, where the same backbone conditioned on a goal image or language, and published demonstration-conditioned methods, complete at most 19. On three of Meta-World’s five held-out classes it beats the best published result, though not on the five-class mean. The same program runs on four robot arms without retraining. After a push, re-solving completes 23 of 60 episodes where replaying the demonstration completes 6. On a Franka arm fine-tuned on real demonstrations of other tasks, programs compiled from single human videos complete 36 of 40 trials, against 22 for the same backbone given the video’s last frame as a goal image. Because a program is a set of coordinates, a person can edit it, and the authors report that an edited placement follows a shifted terminal state. The abstract page lists no code repository.

A planner between the world model and the action (ViGAR)

ViGAR accepts the frame and argues the weakness is horizon. Existing models jointly predict short-horizon visual futures and actions, the authors write, but lack explicit subtask-level reasoning. Their fix factorizes manipulation into a visual subgoal planner and a subgoal executor that share a pretrained world-model representation, so planning and execution draw on the same physical knowledge. The benefit claimed beyond accuracy is in-context learning: a global goal image as context induces different subtask decompositions with no parameter updates. On RoboTwin Clean2Random the authors report 82.00 percent success in Clean and 67.02 percent in Random, which they describe as 12.86 percentage points above the strongest baseline in average success, plus real-robot experiments on five compositional tasks and two in-context learning tasks. No repository appears on the abstract page.

The generative source (ProAct)

ProAct targets a different object: not what the model knows but where it starts sampling. Flow-based VLA policies, it observes, transport samples from a task-agnostic isotropic Gaussian source conditioned on neither recent execution nor predicted future evolution. That discards the continuity of the motion just executed, and reduces a world representation to steering the transport rather than deciding where generation begins. ProAct’s Proposal Expert converts recent actions into a scene-aware hypothesis through one motion-anchored endpoint flow-matching step, initializing generation near the demonstrated action manifold. A World Expert predicts a task-consistent latent future, and the two calibrate a proposal-centered anisotropic source that bounds per-step deviation across translation, rotation and gripper directions. Against the pi-0.5 baseline the authors report better performance across simulation and real-world tasks while reducing denoising steps by 50 percent, inference latency by up to 25.8 percent and raising throughput by up to 34.8 percent.

Fix the data instead (NEEDLEWORK and Awomo)

NEEDLEWORK starts from the observation that demonstrations may contain useful behaviour even when episodes are inefficient or unsuccessful, and stitches short, verified action bridges between recorded observations. The constraints are the interesting part: only RGB images, proprioception and episode-level outcomes, with no new environment interaction and no privileged object state, and the original demonstrations stay in training. On real-robot tasks the authors report an average improvement of 21 percentage points over the strongest baseline on each task. The paper states it was submitted to ICLR 2027.

Awomo goes further upstream and generates the worlds that demonstrations are collected in: asset services for rigid and articulated objects with structure-grounded parts and joints, scene generation that reconstructs editable scenes from images or builds rooms from text, a harness that validates scenes and routes failures back to the responsible module, and PolicyForge, which binds validated worlds to tasks and embodiments to emit replayable demonstrations. On MuJoCo-based LIBERO-Plus, co-training with Isaac Sim demonstrations improves a world action model’s overall success rate from 77.17 percent to 89.43 percent, with goal success up 31.66 percentage points and spatial success up 6.25, and the authors state that gains are more limited on long-horizon tasks.

The test only one of them ran

Generality claims are usually measured on held-out episodes of the same task family. PEWAM reports two tests that are not: one program on four arms without retraining, and closed-loop re-solving after a push against trajectory replay. Both describe a property of the representation rather than the parameter count. Runway’s cross-embodiment evidence is still a partner list, ViGAR’s in-context result sits inside one benchmark, and ProAct’s most transferable claim is control cost: halving denoising steps at a 20 Hz control loop decides whether a policy runs on the robot or beside it.

Why an average cannot settle the argument

Fig, a company building agent foundation models, published an item-level evaluation on 29 September that ran seven frontier models over 177 VisualWebArena tasks, one pass per cell, with a plus or minus 3 percentage point noise band, and three models over four offline physical domains including VLABench manipulation and Bench2Drive driving. Its central finding is that the ranking changes with the axis you sort by, and that in all 21 model pairs the lower-scoring model solves at least three tasks the higher-scoring one fails, up to 21 tasks in the widest pair. The top model solves 139 of 177 tasks, 78.5 percent, while the seven together solve 156, or 88.1 percent. Sixty-six tasks are solved by every model and 21 by none, so the entire 20.9 point spread between best and worst comes from the contested 90 tasks, just over half the set. The same unevenness appears in the physical domains, where models that beat Gemini 3.5-Flash on every domain average still do worse on some task categories. The item-level results were released as the RIDGE dataset.

Read that against this week’s numbers. 82.00 percent on RoboTwin Clean, 89.43 percent on LIBERO-Plus and 40 of 90 episodes on LIBERO-Goal are means over task sets their authors chose and labelled, and with one pass per cell the per-task flips stay invisible. A mean that improves by 12.86 points can coexist with new failures on tasks the previous best system solved. The useful artefact is the per-task table, and almost nobody ships it.

What this means if you are building a robot

• Hold out an embodiment, not just episodes. Held-out episodes on the same arm or benchmark family do not show what PEWAM’s four-arm transfer shows. Ask what the policy conditioned on when the arm changed.
• Ask for the perturbation number. Closed-loop re-solving against replay is the mechanism behind 23 of 60 after a push against 6 of 60, and an unexpected state change is what your robot will actually meet.
• Treat efficiency claims as the deployment claim. Fifty percent fewer denoising steps and up to 34.8 percent more throughput decide whether a flow policy fits your control loop.
• Read the repository, not the abstract. As of today the PEWAM, ViGAR and ProAct pages list no code. For the architecture background behind these policies, see why vision-language-action models matter for robotics.

What to watch next

Whether Runway’s open-weight release arrives with an independent check on the 0.95 correlation, since that number carries most of the video-pretraining argument. Whether cross-embodiment transfer becomes a standard column in results tables, the way scaling curves did for language models. Whether anyone publishes per-task flips alongside means, since a rising average with a shrinking union of solved tasks would be a real finding rather than a marketing problem.

Sources

• PEWAM, Keep the Effect, Drop the Actor: Programmable Effect-to-Execution World-Action Models, arXiv:2610.02398, https://arxiv.org/abs/2610.02398
• ViGAR, Rethinking World-Action Model for Compositional and In-Context Robotic Manipulation, arXiv:2610.02368, https://arxiv.org/abs/2610.02368
• ProAct, World-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models, arXiv:2610.02323, https://arxiv.org/abs/2610.02323
• NEEDLEWORK, Offline Rewriting of Robot Data with Verified Local Stitches, arXiv:2610.02339, https://arxiv.org/abs/2610.02339
• Awomo-SimDataEngine, Agentic Simulation-Ready World Generation, arXiv:2610.02274, https://arxiv.org/abs/2610.02274
• Runway, Introducing Praxis-1, https://runwayml.com/research/introducing-praxis-1
• The Robot Report, Runway introduces Praxis-1 world action model for robotics, https://www.therobotreport.com/runway-introduces-praxis-1-world-action-model-robotics/
• Fig, Astra, Opus 5.5, and other Frontier Models Demonstrate Jagged Performance Across SoTA Agentic Tasks from Web Browsing to Robotics, https://www.fig.inc/blog/astra-opus-5-5-and-other-frontier-models-demonstrate-jagged-performance-across-sota-agentic-tasks-from-web-browsing-to-robotics/