A robot moving a cup from a table toward the sink must answer two questions: how will the scene evolve next, and what action should I take to steer the scene toward my goal? The first belongs to video generation models, the second to robot policies. These have mostly been handled separately—or crammed into one network in ways no one could attribute—which is the problem OpenWAM attacks.
On September 7, a 24-member joint team from seven universities including the National University of Singapore, Tsinghua University, Peking University, and Shanghai Jiao Tong University released 《OpenWAM: An Open, Modular Exploration Towards Systematic World–Action Model Pretraining》 (arXiv:2609.07398), decomposing this route into replaceable parts and delivering a list of design rules derived from controlled experiments. Code, weights, and data recipes are open-sourced.
🧱 Three Layers: Infra, Study, and alpha
OpenWAM-Infra is modular infrastructure splitting four layers—model, training runtime, deployment runtime, and evaluation protocol—behind explicit interfaces. The model layer is assembled from a vision encoder, a swappable flow backbone, and a visibility attention mask. A unified runtime handles pretraining, fine-tuning, and checkpoint recovery, saving full configs, module reconstruction info, and action normalization statistics. Deployment uses a single policy server supporting sync/async inference and multiple denoising schedules. The evaluation side uses a lightweight client with identical I/O for simulation and real robots.
OpenWAM-Study is a set of controlled experiments on RoboTwin2.0 under fixed training budgets and evaluation protocols, answering three questions: which world knowledge to inherit, how world modeling and action learning should coordinate, and how that coordination scales across data domains and embodiments.
OpenWAM-alpha instantiates the Study conclusions: ~6,400 hours of first-person human and robot data, an 80-dimensional unified action representation, and a dual-system architecture.
🔩 Three Architecture Families
Infra supports three assembly styles. The single-system design shares one backbone for video and action. The dual-system design uses two backbones connected via joint self-attention, cross-attention, or inverse dynamics. The triple-system adds a vision-language understanding stream. All are assembled via unified configs and a registry, so trainers and the policy server need not know which is loaded.
The point is variable control: swap backbone, encoder, or information flow while everything else stays fixed. Before this, world-action models across papers differed even in skeleton, making result differences unattributable.
📐 Rule One: Strong Priors, Compact Latents
Comparing four video backbones by average success rate: 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%. Bigger capacity wins, but the team chose 5B as default—it trails 14B by only 1.40% at much lower cost. The paper itself cautions that these four models differ in architecture, data, and training objectives, so gains cannot be attributed to parameter count alone.
On visual representations, the key is not whether the encoder is reconstruction- or representation-style, but whether the latent space is compact across both time and token dimensions while retaining sufficient world information. DINOv3 and V-JEPA 2.1, producing high-dimensional features directly, performed worse initially; compressing them via S-VAE into DiT-compatible spaces lifted performance significantly, with DINOv3 + S-VAE approaching Wan2.2-VAE.
| Video backbone | Avg success rate | Note | |---|---|---| | Wan2.1-VACE-1.3B | 90.14% | Smallest capacity | | Cosmos-Predict2.5-2B | 91.64% | NVIDIA family | | Wan2.2-TI2V-5B | 92.39% | Chosen as default | | Wan2.1-I2V-14B | 93.79% | Best, most expensive |
🔁 Rule Two: Build World-to-Action Information Flow Explicitly
Architecture ablations show average success rate rising stepwise from single- to dual- to triple-system, driven by an explicit world-to-action information flow established during training; inference uses synchronous joint denoising so future prediction and action generation constrain each other in step. This converts the vague debate over "whether world models help" into tunable parameters about which layer feeds which.
🧪 Eight Benchmarks, Four Embodiments
The evaluation protocol covers eight simulation benchmarks: LIBERO, LIBERO-Plus, VLABench, RoboTwin2.0, RoboDojo, RoboCasa365, EBench, RoboCasa-GR1, spanning single-arm, bimanual, mobile manipulation, and dexterous hands. Open-sourced content includes infrastructure, evaluation protocols, pretraining and fine-tuning weights, and data recipes.
📼 Data Asymmetry and the Road Here
The route's unavoidable reality: internet video is measured in hundreds of millions of hours while robot trajectories are measured in tens of thousands. Inheriting visual dynamics priors from video generation pretraining, then learning how to act from embodied experience, is the core bet. From UniPi's "policy as video" onward, the line has passed through decoupling, alignment, joint modeling, and large-scale pretraining; by 2026 the focus has shifted from "can we predict the future" to "which future is actually useful for action."
🔭 Things to Watch
- Third-party replication: whether independent teams reproduce Study's conclusions with OpenWAM-Infra.
- Backbone upgrade curve: whether the 90.14%–93.79% band shifts wholesale with next-gen video models.
- Cross-embodiment cost: information loss of the 80-dim unified action space on dexterous hands.
- Real-robot data ratio: the split between first-person human and robot data within 6,400 hours is unpublished.
- Triple-system cost-effectiveness: whether the extra vision-language stream justifies its inference overhead.
References
1. OpenWAM team, *OpenWAM: An Open, Modular Exploration Towards Systematic World–Action Model Pretraining*, arXiv:2609.07398 2. Project page openwam-official.github.io; code github.com/OpenWAM-Official/OpenWAM; models and data at huggingface.co/OpenWAM 3. Synced Pro coverage, 2026-09-25 4. Team interview on embodied model design insights, 2026-09 5. BestHub technical teardown of OpenWAM-Infra and OpenWAM-Study, 2026-09-25