Seven universities open-source OpenWAM: a full-stack world action model framework and base weights
According to Machine Heart, a team from seven universities including the National University of Singapore, Tsinghua University, and Peking University has open-sourced OpenWAM, comprising the modular infrastructure OpenWAM-Infra, the controlled-experiment study OpenWAM-Study, and the base model OpenWAM-α, covering eight simulation benchmarks and three classes of real robot platforms, with pretrained weights and a data recipe released.
What changed
According to Machine Heart, a research team from seven universities including the National University of Singapore, Tsinghua University, and Peking University has open-sourced OpenWAM, positioned as a full-stack research framework and base model for world action models (WAM). The project has three layers. OpenWAM-Infra provides composable models (the vision encoder, data-flow backbone, and visibility attention mask can be swapped independently, supporting single-system, dual-system, and tri-system architectures), a unified training runtime (pretraining, fine-tuning, and checkpoint resumption, with checkpoints saving the full configuration and action normalization statistics), a deployment runtime (the same policy server hosts different models, supporting synchronous and asynchronous inference and multiple denoising schedules), and an evaluation protocol (covering eight simulation benchmarks: LIBERO, LIBERO-Plus, VLABench, RoboTwin2.0, RoboDojo, RoboCasa365, EBench, and RoboCasa-GR1). OpenWAM-Study conducts controlled experiments under a fixed budget, reporting backbone capacity comparisons (Wan2.1-VACE-1.3B 90.14%, Cosmos-Predict2.5-2B 91.64%, Wan2.2-TI2V-5B 92.39%, Wan2.1-I2V-14B 93.79%), information-flow ablations (action stream attending one-way to the world stream 92.39%, bidirectional mutual attention 92.15%), and cross-domain embodied pretraining results (OOD success rate 14.50% from scratch, 26.62% with co-training). OpenWAM-α uses Wan2.2-VAE as a frozen vision encoder, Wan2.2-TI2V-5B as the video-stream backbone, and a 1B-parameter ActionDiT for the action stream, with an 80-dimensional unified action space, and the pretraining set is about 518.5 million frames, roughly 6,369 hours. The paper states that it leads the second place by about 4 percentage points in both SR and Score on EBench, reaches an average success rate of 82.5% across six single-arm real-robot tasks, and reports results after fine-tuning on an unseen dexterous-hand morphology.
What it means for developers
For developers working on robot learning, the impact is mainly in three areas, all inferences from the source report rather than verified facts. First, the unified action space plus validity mask approach maps joint/end-effector actions from different platforms into fixed slots, so cross-morphology fine-tuning does not require rewriting the action head; this lowers the integration cost of adding a new robot platform, but at the price that you must first define the mapping convention from actions and states to slots. Second, the deployment data reported by the source (synchronous inference, 10-step joint denoising, and, via fixed-shape compilation, CUDA Graph, velocity caching, prompt embedding caching, and skipping VAE decoding on the control path, about 170 ms per action chunk on an RTX 5090) indicates that its inference stack is heavily optimized for fixed shapes and caching; if your input resolution, action chunk length, or batch size changes frequently, the benefit of such optimizations will decay noticeably. Third, the source explicitly points out the division of labor between WAM and VLA: WAM fits the training distribution better on in-distribution tasks, while VLA is usually more robust under out-of-distribution perturbations; the degradation on LIBERO-Plus is concentrated in camera and noise perturbations, attributed to the limited proportion of similar single-arm perturbation samples in the pretraining data and to pixel-latent long-horizon prediction being more sensitive to viewpoint noise. This means that applying WAM directly to sites with varying sensor layouts or high-frequency jitter requires targeted data reinforcement first, rather than expecting out-of-the-box generalization.
Integration and evaluation
The following are recommendations and conditional analysis, not facts stated by the source. Recommendation 1: whether to replace the video backbone depends on your compute budget and tolerance for success-rate differences. The source reports that the 14B backbone is 1.40 percentage points above the 5B, while the 5B was chosen as the default, so if you are in a stage with a single machine and multiple GPUs and need frequent retraining, start with the 5B as a baseline and set your evaluation criterion as an average success rate no lower than your current approach under the same RoboTwin2.0 configuration and fixed budget, while recording single-training and per-action-chunk inference latency. If your task is high-precision long-horizon manipulation and you already have stable compute, then verify whether the 14B brings gains above the noise level. Recommendation 2: whether to adopt the bidirectional mutual-attention information flow depends on whether you have done cross-domain embodied pretraining. The source reports that from scratch the one-way world-to-action flow has a slight advantage, while after pretraining the bidirectional mutual attention is more stable in-domain and out-of-domain, so if you directly fine-tune the official pretrained weights, prioritize reproducing the bidirectional mutual-attention configuration; if you train a small-data model from scratch, compare the one-way setting first. A suggested evaluation criterion is the success rate under Clean2Random-style randomization, with the requirement that in-domain metrics do not drop noticeably below the baseline. Recommendation 3: whether to invest in human first-person-view data depends on whether your bottleneck is in-domain accuracy or out-of-distribution generalization. The source reports that pretraining gives limited in-domain improvement but significant OOD improvement, and that single-stage co-training is slightly better than staged training. So if your site mainly involves new lighting, new backgrounds, and new camera poses, prioritize expanding mixed human first-person-view and robot data and do single-stage training; if your problem is a high basic grasping failure rate, first improve the action annotation quality of robot trajectories. Recommendation 4: before integrating the action space, use a small-scale replay to verify that unused dimensions in the 80-dimensional mapping are correctly masked by the validity mask, then evaluate post-fine-tuning task success rate, to avoid invalid dimensions polluting the training signal.
Limits and open questions
OpenWAM-α does not currently appear in KimiSeek's live model catalog, so developers cannot call it directly on this site; the model and data releases mentioned in the source point to project pages such as Hugging Face, which is unrelated to availability on this site. All performance numbers in the source (including average success rates of each backbone, the EBench lead, and real-robot success rates) come from the paper authors' own claims; this article has not independently reproduced them, nor re-compared these methods under unified conditions; the different backbones are not entirely identical in structure, data, and objectives, and the source itself cautions against simply attributing gains to parameter count. The real-robot results come from the research team's experimental platform, and the task set, evaluation protocol, and hardware configuration differ substantially from production sites and cannot be directly extrapolated. Expressions such as "first tier" in the source lack complete per-benchmark rankings, so readers should rely on the paper's panel data. Whether the project uses an open-source license and whether the weights permit commercial use are not stated in the source; downloadable weights do not equal an open-source license, and the actual terms of the repository and model card should govern. Reproducibility depends on whether the pretrained weights, data recipe, and evaluation protocol can be fully obtained; the source says they have been released, but the actual completeness and future maintenance strategy remain unknown; the distribution license of the data sources may also restrict your retraining use.