Embodied AI company Anyverse Dynamics has released and open-sourced MWA-Phoenix, a latent-space world action model meant to serve as a foundation model for robot manipulation. The pretrained model has 7 billion parameters and is designed to run on the robot itself. Code is on GitHub and a technical report has been published alongside it.
MWA-Phoenix works from RGB camera input only and does not use ground-truth depth. According to the company, it ranked first or near the top on six evaluations with different emphases: 94.63% on LIBERO-Plus, which adds disturbances such as camera angle, lighting and sensor noise; 66.48% on LIBERO-Pro, which changes objects, positions and goals; 82.67% on RoboCasa GR1 Tabletop, up 7.5 points from its previous WALA model; and 72.42% on the RoboTwin 2.0 dual-arm benchmark. It also reports results on the RoboDojo real-robot leaderboard, which covers three robot bodies and 18 tasks.
The most telling number may be on Bench2Dex, which spans 12 combinations of robot arms and dexterous hands. Phoenix averaged 45.6%, which the company says is 17.3 points ahead of the strongest baseline, pi0.5. The absolute score shows that transferring skills across different robot bodies remains far from solved.
Technically, the model uses DINOv3 as its vision backbone and distills knowledge from two teachers: VGGT for geometry and V-JEPA for temporal change. A causal semantic VAE compresses those features into a compact representation without reading future frames. A video expert and an action expert are then linked through joint attention, so the model predicts how the scene will change and what the robot should do at the same time.
Two companion releases target the gap between simulation and real hardware. Phoenix-ILA folds motion and contact into action learning for fine manipulation, while Phoenix-PhysCanvas is an interactive simulated world model where policies can keep improving in the cloud before deployment. For real-time control, the team uses asynchronous execution so the next action chunk is prepared while the robot is still moving.
Pretraining data comes from real robot operation, egocentric human video and handheld UMI capture. Anyverse Dynamics says it holds more than 200,000 hours of multimodal data and is training larger models. All benchmark figures are the company's own, and it says no test data was replicated in training.
SEE ALSO: Qianjue Unveils Polibrain Predictive World Model · Shengshu Technology Releases Motus2