Robotic Hands
Robotic hands still can't match human dexterity, partly because training them directly on real hardware is slow, expensive, and rough on the equipment. Instead of training on real hardware, I trained a control policy in simulation. A major challenge with this is that simulated physics never quite matches reality, so behaviour that works perfectly in simulation often breaks down on real hardware.
So, I measured the physical properties of a single joint on a robotic hand:
- Friction
- Gearbox behaviour
- etc.
and used those measurements to calibrate a physics simulation.
I then trained a reinforcement learning policy in that simulation to move the joint to target positions using torque control. Once trained, I deployed the policy straight onto the physical joint with no real-world data and no fine-tuning.
It worked: the policy controlled the real joint successfully despite never having seen it during training.
I followed this up with an ablation study to work out which parts of the calibration actually mattered for the transfer to succeed.
Source code, results, etc. are unfortunately not publicly available.