Dexterous in-hand manipulation usually relies on vision, but humans can rotate objects using touch alone. Robots have struggled to match this, especially when moving from simulation to real hardware. Researchers trained a reinforcement learning policy for Baoding ball rotation on the Shadow Dexterous Hand Lite, using only proprioception and binary tactile feedback as observations.
To make the policy transfer without retraining, the team aligned the simulated hand's sensors and coupling with the physical hand, calibrated the tactile sensors, and introduced two novel domain randomisations: slew-rate action limits and stuck-at taxels. This allowed zero-shot transfer to the real robot, which then completed 225 consecutive 180-degree rotations—112.5 full rotations—at roughly 0.36 rotations per second in a single continuous trial, with no vision or object-state information.
The authors report this as the strongest hardware result for Baoding ball manipulation in both rotation count and speed, regardless of sensing modality. They also compared tactile sensing against proprioception alone, identifying which hand regions benefit most from touch. Since this is a single source, there are no conflicting findings to note. The work sets a new benchmark for sustained, tactile-driven dexterous manipulation on physical hardware.