OpenAI Dexterous Hand: fine manipulation with reduced sim-to-real gap
In one sentence OpenAI advances robotic dexterity research with new results on reduced sim-to-real gap via massive domain randomization and modern RL on the Shadow Hand.
In 2019, OpenAI had already shown a robotic hand capable of solving a Rubik's cube. The problem was that it only worked under very precise conditions and often failed in the real world. The new results show significant progress in closing this gap.
The key technique is called domain randomization: instead of training the robot in a fixed, realistic simulation, it trains across thousands of slightly chaotic and varied simulations. The robot learns to be robust to variations, and when it encounters the real physical world, it treats it as "just another different simulation."
The result is a robotic hand that manipulates objects finely and adaptively, even in conditions it never saw exactly during training.
Companies
OpenAI, Shadow Robot
Tools
Shadow Dexterous Hand, Asymmetric Actor-Critic, OpenAI Gym
Tags
Sources