Memo Releases Physical-World Action Model and Evaluation Benchmark for Embodied AI

In September 2026, Hangzhou-based Memo unveiled Physical-WAM and RoboTwin-Phys at the fifth Global Digital Trade Expo. Physical-WAM is described as the world's first physics-aware physical-world action model, while RoboTwin-Phys is introduced as the industry's first benchmark for physical drift evaluation.

Physical-WAM is trained on observable and deeper physics data. It combines three modules—PhysLens, PhysDream, and PhysAct—into a closed loop of perception, prediction, generation, and correction. PhysLens extracts physical cues such as friction, weight, and center of mass from multimodal signals; PhysDream forecasts future physical states and risks like slipping; and PhysAct adjusts action generation based on physical conditions. RoboTwin-Phys adds 13 common real-world physical perturbation types across dimensions including object attributes, contact properties, damping, task environments, and camera setups, enabling both single-factor and combined-factor tests.

Memo says embodied intelligence faces both a shortage of real physical interaction data and a lack of evaluation standards. CEO Li Minghao said during the expo that robots must genuinely understand the physical world before embodied AI can move from labs to large-scale industrial deployment. The company was incubated by the AVS Advanced Visual System Research Center of Peking University's Advanced Institute of Information Technology, and received exclusive seed funding from Suochen Technology in August 2026.