Stanford's Fei-Fei Li Team Releases OpenWAM World-Action Model Framework

The term world-action model (WAM) was explicitly introduced by Nvidia in its DreamZero paper in February this year, and related papers have emerged rapidly since. On September 22, Black Forest Labs, known for the FLUX image model, released the open-source 7B world-action model FLUX 3 Action; according to its official blog, the model surpassed the previously strongest open-source model Cosmos 3 Nano on Nvidia's RoboLab-120 leaderboard, despite having less than half the parameter count. This claim currently comes only from a vendor blog and lacks third-party verification.

The problem is that these systems almost all simultaneously change the video backbone, the training data, the interaction method between video and action, and the inference pipeline. When all variables move together, no one can say clearly which design brought the improvement. The team of Fei-Fei Li, Jiajun Wu, Ehsan Adeli and others at Stanford University released a paper proposing the open world-action modeling framework OpenWAM, attempting to give this question a systematic answer. The paper is titled OpenWAM: An Open Framework for Composable World-Action Models, with the paper at arxiv.org/abs/2610.07922, the project page at openwam.stanford.edu, and the code repository at github.com/OpenWAM/OpenWAM.

OpenWAM starts from Alibaba's open-source video model Wan2.2-5B, first further pretrained on roughly 3.34 million robot and human interaction videos, a nominal duration of about 14,600 hours, with no action labels used throughout, trained for 14 days on 32 B200s. The purpose of this step is to give the model built-in physical priors rather than learning actions from scratch.

The key change is "causalization": each video chunk can only see the current and preceding content, and cannot peek at the future. The original model is like holding the whole script and repeatedly revising lines, whereas now it must perform forward in time like a live broadcast, which is exactly what a robot acting step by step requires. The team then uses a Mixture-of-Transformers (MoT) architecture to splice together a 5B video expert and a 2B action expert, each retaining its own normalization, projection and feed-forward layers, exchanging information only in one layer of joint attention.

On this skeleton, the team defines four interaction procedures: VTA imagines the picture first and then generates actions, ATV does the reverse order, Joint lets both generate simultaneously and reference each other, and Decoupled makes the two mutually invisible in the future portion. All four share the backbone, tokenization and flow matching objective; the only variable is generation order and cross-modal attention, so the comparison takes on the character of a controlled variable.

The more novel part of the paper is splitting the inverse dynamics model (IDM) and the forward dynamics model (FDM) into components that can be trained independently. The IDM is like a translator: it takes the current frame, the robot's proprioceptive state and a segment of imagined future video as input, and outputs concrete actions; the FDM does the reverse, predicting what the picture will become based on actions.

The key lies in "local": neither receives task instructions or earlier history, and both care only about the correspondence between this small segment of picture change and actions. They are therefore not bound to a specific task and can be spliced with any compatible video predictor — the video model decides what should happen, and the IDM is responsible for how to achieve it.

But a translator that has only learned from successful demonstrations has too narrow an experience. To this end, the team built the LIBERO-Long-CF dataset: restoring the simulator states in the demonstrations and executing altered actions, such as stopping, reversing, adding noise, changing the timing of gripper opening and closing, and so on, yielding a total of 32,000 clips and about 4.1 million control records, 29.7 times the original demonstrations, with 72.3% of the clips changing the object's placement state. Many branches simply cannot complete the task, but that is precisely the point.

On the four LIBERO subsets, the success rates of the four interaction procedures are all high, with VTA averaging 98.6%, slightly above reported results cited in the paper such as LingBot-VA (98.5%) and π0.5 (96.9%). However, LIBERO is already near saturation, so the reference value of its success-rate differences is limited. The more noteworthy detail is that on LIBERO-Long, Decoupled (97.0%), in which the future portion is mutually invisible, is on par with the bidirectionally interacting Joint (96.6%).

On real robots, the team used two Franka FR3 robotic arms to test toasting bread, restoring the last layer of a 2×2 Rubik's cube, and sorting cups by color, with VTA and Joint averaging success rates of 92.1% and 91.9% respectively.

The ablation experiments show the weight of the backbone: with other conditions unchanged, when initialized with the original Wan2.2, VTA reaches only 68.4% on LIBERO-Long, rising to 97.8% with the causal backbone pretrained on robot videos, an improvement of 29.4 percentage points. The authors emphasize that this is the combined effect of the data and the causalization change, not a single factor.

The most critical part is the component transfer experiment. The team freezes the IDM trained on LIBERO-Long and pairs it with a video predictor adapted for four new LIBERO-90 tasks. As a result, the local IDM trained on demonstrations plus counterfactual data reaches an average success rate of 84.0%, while the IDM receiving full context gets only 47.0%, and the local IDM trained only on demonstrations gets only 21.5%. "Looking only at the local" and "having seen sufficiently diverse consequences" are both indispensable.

It should be noted that the video predictor on the target task was fine-tuned with demonstrations containing action labels, so this is not zero-shot transfer, and the authors state this explicitly.

The FDM conclusion is consistent. Starting from the same state and executing 16 different actions, letting the model judge which real outcome the predicted picture corresponds to, random guessing is about 6%, the FDM trained only on demonstrations is 21.1%, and after adding counterfactual data it rises to 71.3%, and the model can finally distinguish the difference between "pushing this way" and "pushing that way."

The authors also candidly acknowledge the limitations: component reuse is mainly verified in simulation, and the cost of "deliberately doing it wrong" on real robots is much higher; the FDM also only covers short-horizon prediction. This means it is currently uncertain whether the counterfactual data construction method can be transferred to real robot scenarios at low cost.

In addition, the peer review status and independent reproduction of the OpenWAM paper are not mentioned in the sources, and the actual availability and maintenance of the open-source code and pretrained video model weights are also not explained. Against the backdrop of systems in the WAM field simultaneously changing the video backbone, training data, interaction method and inference pipeline, the contribution of a single design to performance improvement remains difficult to fully attribute — OpenWAM provides a controlled testbed, not the final answer to the attribution problem.