Robots can now learn to handle objects by studying the physics of how humans touch them, not just where their hands go.
Researchers have released CHORD, short for Contact Wrench Guidance from Human Demonstration in Robotic Dexterous Manipulation. Instead of teaching a robot to mimic the shape of a human movement, CHORD translates demonstrations into the forces and torques a hand exerts on an object — what engineers call a contact wrench. A reinforcement learning system then trains against that physical signature. The team tested the approach on 1,831 tasks drawn from a new benchmark of 4,739 bimanual manipulation scenarios built from motion-capture data and reconstructed video. CHORD hit an average success rate of 82.12% across those tasks and 90.77% when generalizing from hand-only or third-person video to full-body robot control.
The force-centric framing matters because it sidesteps a long-standing gap between human and robot anatomy. A five-fingered human hand and a robotic gripper move differently, so motion imitation tends to break down on contact-rich tasks like opening jars or manipulating hinged objects. By anchoring the comparison to what forces land on the object rather than how the limb moves, CHORD gives the learning signal a geometry-independent anchor. The policies also transferred to real hardware in both open-loop and closed-loop settings, which is where most lab results quietly die.
Dexterous manipulation has been robotics' stubborn last mile for years — Tesla, Figure, and a wave of humanoid startups have all made it a centerpiece pitch. A benchmark of nearly 5,000 tasks and results that survive contact with actual hardware is a more credible data point than most demo reels offer, though peer review and third-party replication are still ahead.