Gasgoo Munich- Anyverse Dynamics has released and open-sourced MWA™-Phoenix, a fully pre-trained, 7-billion-parameter latent-space world model foundation built for edge deployment.
Public benchmarks show the model topping six simulation and real-robot leaderboards — including LIBERO-Plus, LIBERO-Pro, and the RoboDojo real-robot ranking. It hit a 94.63% overall success rate on LIBERO-Plus. Across all real-robot demonstrations released this time, Xense Robotics's photon tactile sensors served as the tactile input source, feeding real contact data directly into action learning and interaction prediction.
Image source: Xense Robotics Robotics
On the action-learning front, Anyverse Dynamics' Phoenix-ILA module encodes synchronized motion and tactile sequences into a shared interactive latent space, jointly learning movement, local tactile deformation, and force-related signals. Ablation studies reveal that adding this module doubled the success rate for cucumber peeling — from 20% to 40% — and lifted plate-wiping accuracy from 75% to 90%. For tasks demanding continuous surface contact, real-world interaction data gives action learning a direct grounding.
During pre-deployment policy training, Anyverse Dynamics' in-house Phoenix-PhysCanvas visual-tactile world model predicts the visual and tactile feedback after an action is executed, supporting reinforcement-learning post-training. Training runs in an interactive simulation environment before results are validated on real robots. Experiments show average real-robot success rates climbing from 58.33% to 75.00% for test-tube tasks, and from 80.00% to 86.67% for block-stacking tasks — gains of 16.67 and 6.67 percentage points, respectively.
The real-robot tasks — twisting bottle caps, peeling, wiping, and long-horizon operations — all rely on Xense Robotics sensors for signal capture. To safeguard data quality, Xense Robotics built cross-unit consistency and repeatability testing standards into its sensor R&D and manufacturing, strictly aligning the temporal relationships among visual, tactile, and action signals. Meanwhile, tactile simulation and modeling continuously verify the data's physical consistency, ensuring recorded contact information obeys real-world physics and can be directly used as training samples.
Following the model's open-source release and real-world application, Xense Robotics says it will keep refining its tactile hardware and data infrastructure. Field testing by frontier teams has yielded crucial product feedback, clarifying which contact details matter and how perception should adapt to specific tasks. Going forward, Xense Robotics plans to fuel the steady evolution of robotic manipulation by supplying high-quality, real-world tactile data.
