Dactyl

OpenAI's project that trained a five-fingered robot hand entirely in simulation with reinforcement learning, then transferred it directly to the real hand.

Dactyl is OpenAI's dexterous manipulation project, built on the Shadow Dexterous Hand, a five-fingered robotic hand. The 2018 paper 'Learning Dexterous In-Hand Manipulation' had the hand reorient a block to a target pose in its palm: the policy was trained entirely in simulation with reinforcement learning and transferred straight to the real hand using domain randomization (randomizing physical parameters such as friction and object appearance during training), with no human demonstrations at all; human-like behaviors such as finger gaiting emerged on their own. The 2019 follow-up, 'Solving Rubik's Cube with a Robot Hand,' introduced automatic domain randomization (ADR), which automatically widens the randomization ranges as training progresses. A Kociemba solver computes the cube-solving move sequence; the neural network only handles the physical manipulation. The paper reports about 60% success at 15 face rotations and about 20% at the hardest 26-rotation sequences. Dactyl is a landmark result for both sim-to-real transfer and dexterous manipulation.

ExampleIn the Rubik's Cube experiments, the cube's pose was estimated from three camera views with a CNN, individual face angles were read from a sensor-equipped 'Giiker' smart cube, and fingertip positions came from a motion-capture system — all fed into a recurrent neural network policy that controlled the fingers.

Also called

OpenAI Dactyl, Learning Dexterous In-Hand Manipulation, Solving Rubik's Cube with a Robot Hand

Related

Automatic Domain Randomization · Domain Randomization · Sim-to-Real Transfer · In-hand Manipulation · Shadow Dexterous Hand · Proximal Policy Optimization

Sources

arXiv 1808.00177: Learning Dexterous In-Hand Manipulation
arXiv 1910.07113: Solving Rubik's Cube with a Robot Hand

As of

2019-10