DexDrummer: Sim-to-Real Robot Drumming via Hierarchical RL with 20-DOF Dexterous Hands

arXiv / Import AI 451·low signal

Researchers published DexDrummer (arXiv 2603.22263), demonstrating hierarchical reinforcement learning for in-hand, contact-rich dexterous manipulation on real Franka Panda arms with Tesollo DG-5F hands (20 degrees of freedom). The dexterous reactive policy outperforms fixed-grasp baselines by 1.87x on easy songs and achieves F1 scores of 1.0 on training songs with zero-shot sim-to-real transfer. As Import AI notes, the limited competency on complex songs demonstrates how far real-world robotics lags behind other AI domains.