AdaDex: Hybrid Dexterous Mimicking for Adaptable Robotic Skills
AdaDex is a goal-conditioned reinforcement learning framework that enables robots to learn adaptable, human-like dexterous skills from minimal human demonstration. It addresses the trade-off between natural motion tracking and robust generalization by combining both approaches.
Reading in Plain English mode — toggle below to switch
AdaDex is a goal-conditioned reinforcement learning framework that enables robots to learn adaptable, human-like dexterous skills from minimal human demonstration. It addresses the trade-off between natural motion tracking and robust generalization by combining both approaches.
Imagine learning to juggle. A coach (motion tracking) shows you exactly how to move your hands, ensuring your form is correct. But to truly master it, you also need to practice on your own, even when you drop balls (generalization environments), learning to recover and adapt to different throws. AdaDex combines both: precise imitation and robust self-correction.
How AdaDex Works
- 11. **Human Trajectory Input:** AdaDex starts with a short human hand-object trajectory, either from motion capture or reconstructed from video.
- 22. **Hybrid Environment Training:** A single policy is trained simultaneously in two types of parallel environments: motion tracking environments that use a dense objective to mimic human motion, and generalization environments with varied initial states and sparse success rewards.
- 33. **Policy Learning:** The policy learns to balance the fidelity of human-like movements with the ability to achieve goals even when conditions deviate from the original demonstration.
- 44. **Skill Adaptation:** The resulting policy can perform diverse, adaptable dexterous skills like grasping, reorientation, and dynamic throw-and-catch, generalizing across different scenarios and even to the real world.
Lₜotal = L_motion + λ * R_success
This formula represents the total loss function for AdaDex. $L_total$ is the combined objective. $L_motion$ is the dense motion tracking objective, which encourages the robot to mimic human movements. $R_success$ is the sparse object-success reward from generalization environments, pushing the robot to achieve the task goal. $λ$ (lambda) is a weighting factor that balances these two objectives.
AdaDex overcomes the limitations of pure motion tracking (lack of adaptability) and pure goal-reaching RL (unnatural motions) by synergistically combining both, enabling robots to learn adaptable, human-like dexterous skills from minimal demonstrations.