A robot can feel contact forces without any force sensors by estimating them from motion.

Robots that roll, step, and grasp things need to handle touch and push against objects. Standard vision-language-action models turn pictures and text into movements, but they cannot sense physical contact. Adding physical force sensors to a robot is expensive and difficult to install. Meanwhile, basic whole-body controllers keep the robot stable, but they cannot tell useful work forces apart from outside disturbances. To solve this, researchers built FWBC-VLA for wheeled-legged robots. It works without extra force sensors. Instead, an estimator calculates contact strength and changes in force over time. The system turns these contact details into tokens so the high-level policy can detect when contact starts, stays, and stops. A compensation generator then blends these touch estimates and robot body signals into corrective motions for the whole-body controller. In real-world tests wiping whiteboards and opening doors with closers, this framework successfully guided contact-heavy tasks.

Why it matters

This helps robots that need to touch and move objects, like in manufacturing or home assistance, without expensive sensors. It reduces hardware costs and makes robots more reliable in physical tasks like wiping or opening doors.

Questions answered