Can a robot hand feel what it's holding just by listening to its own joints?

In-hand manipulation typically requires vision or tactile sensors to track what the robot is doing. This work shows that joint encoders alone—the simplest, cheapest sensors already on any robotic hand—can work nearly as well. Researchers trained a Transformer to extract object state from the temporal patterns of joint position and velocity readings on a tendon-driven hand, then tested it on real hardware rotating a cube. The approach achieved 3.1× faster rotation speed than baselines and estimated object position 23% more accurately than simpler neural networks.