Solving Rubik's Cube with a Robot Hand
Problem
Solving a Rubik's cube with a real humanoid robot hand requires robust control and state estimation despite the sim-to-real gap.
Why it matters
Real-world dexterous manipulation of complex objects is a long-standing robotics challenge.
What the AI did
Automatic domain randomization (ADR) generated increasingly difficult simulated environments; control policies and vision estimators trained in simulation transferred to the real robot, showing emergent meta-learning.
Results
Solved a Rubik's cube with a humanoid robot hand; no quantitative success rate stated in abstract.
Impact
Demonstrated sim2real transfer solving manipulation of unprecedented complexity on a real robot.
Record
| Venue | arXiv (Cornell University) |
|---|---|
| Published | 2019 |
| Work type | preprint |
| Authors | OpenAI, Ilge Akkaya, Marcin Andrychowicz, Maciek Chociej, Mateusz Litwin, Bob McGrew, Arthur J Petron, Alex Paino et al. |
| Citations | 569 (OpenAlex, at import time) |
| References | 111 |
| Model or method | memory-augmented models trained with automatic domain randomization (ADR) |
| Organization | OpenAI (per abstract URL) |
| Primary source | https://doi.org/10.48550/arxiv.1910.07113 |
| Open access | green |
| Keywords | robot hand, Rubik's cube, domain randomization, sim2real transfer, meta-learning |
Importance rating
significant — thousands to millions affected
Abstract calls it a manipulation problem of unprecedented complexity solved on a real robot; a genuinely new capability in sim2real dexterous robotics, though no scale of beneficiaries is given.