Robot Hands for Modern AI and Real Work
8 hours ago
- The new Atlas hand features 13 degrees of freedom (DOF) with direct actuation, designed for high-fidelity simulation and sim-to-real reinforcement learning.
- The hand has four fingers (including a more dexterous thumb) and no pinky—a design choice based on functionality versus complexity, as adding a pinky would increase cost and volume without significant benefit.
- Key design tradeoffs include balancing dexterity, strength, ruggedness, cost, repairability, and sensing; the hand is similar in size to a large human hand but stronger, capable of carrying over 100 lbs.
- Backdrivability and proprioception enable agile dexterous behaviors, with dense pressure tactile sensors on fingertips and palm for detecting small contact signals.
- Direct human imitation via wearables has limitations for dexterous manipulation; the hand is optimized for RL in simulation to achieve robust control policies with domain randomization.
- The ultimate goal is a reliable hand capable of using tools, turning humanoids into general-purpose machines for physically demanding tasks.