PhyPush: Single push estimates mass and friction without sensors

PhyPush, a physics-guided Transformer framework developed by researchers including Koyo Fujii, Luis Figueredo, and colleagues, addresses a key challenge in robotics: estimating object mass and friction without specialized sensors. Most existing methods rely on force/torque sensors, tactile arrays, or multi-camera motion capture. PhyPush instead uses only the end-effector velocity from a single push—data available on any standard robotic arm. The model incorporates physical constraints from Newton's second law and the Coulomb friction model via a physics-guided loss function, improving physical consistency and generalization to unseen objects and surfaces.

In simulation experiments, PhyPush achieved more than 10% lower error in mass and friction estimation compared to a baseline that had privileged access to full force information. Real-world tests confirmed that PhyPush outperforms purely data-driven loss approaches, demonstrating that physics-guided learning can replace expensive hardware. The work, submitted to the 2026 IEEE/RSJ IROS conference, opens the door for scalable, low-cost robotic manipulation in domains like logistics, manufacturing, and assistive robotics where sensor costs are prohibitive.