Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand

Researchers trained an anthropomorphic robotic hand via reinforcement learning to crawl, steer, and recover from falls using its fingers as legs.

The paper demonstrates a self-contained robotic hand with onboard power and computation that reuses the same fingers for locomotion, body support, and environment interaction while retaining the finger design and position controller. Reinforcement learning training in a simulator calibrated from hardware measurements accounts for the hand's unequal fingers, and in simulation the hand moves faster than with tuned rewards originally designed for quadrupeds. On hardware, task-specific policies enable untethered crawling, steering, and fall recovery, plus keyboard command execution without vision and object pushing guided by overhead visual feedback.

• RL policies enable untethered crawling, steering, and fall recovery on hardware.
• Simulator calibrated from hardware measurements; custom reward formulation beats tuned quadruped rewards.
• Hand executes keyboard commands without vision and pushes objects using overhead visual feedback.
• Onboard power and computation make the platform fully self-contained.