Solving Rubik's Cube with a Robot Hand
In my talk, I would like to introduce how the OpenAI Robotics team solved Rubik's cube with a single human-like robot hand. Despite our model being trained only in simulation, it can be used to solve a very complicated manipulation problem in the real world. To close the sim-to-real gap, we proposed and used a novel algorithm, called automatic domain randomization (ADR), to generates a distribution over randomized environments of ever-increasing difficulty. Both control policies and vision state estimators trained with ADR exhibit vastly improved sim2real transfer. Videos summarizing our results are available: https://openai.com/blog/solving-rubiks-cube/
Lilian Weng is working in the OpenAI Robotics team, with the goal to develop the algorithms to power future general-purpose robots of the future --- that is, a robot that can operate in the complex environment of the real world and carry out most tasks that humans can do. Her daily job involves writing good code, experimenting with new ideas, reading papers, hacking hardware and working with physical robots. Lilian also has an ML tech blog as she believes the best way to learn is by explaining new concepts clearly to others.