Physical AI Lab // Episode 0.4 // One-pager
Ten thousand tries in a virtual world
What simulation is, and why robots can't do without it
Three key points
- Simulation is a "fake world" in a computer, calculated by the laws of physics: push the blocks, and it works out what happens step by step (a physics engine).
- The upside: safe, fast, and many can practise together. In one paper, thousands of robots trained in parallel, and a four-legged robot learned to walk on flat ground in under 4 minutes (R); with simulation moved onto graphics cards, training ran a hundred to a thousand times faster than the conventional pipeline of CPU simulation plus GPU neural networks (technical report, R).
- The downside: simulation isn't quite the same as reality. A policy trained there often needs tuning again on a real robot (the sim-to-real gap).
Evidence: O official · R research · M media
| R | Thousands | of robots training in simulation at once, on one workstation graphics card (paper) |
| R | Under four minutes | for the four-legged robot ANYmal to learn to walk on flat ground (paper, massively parallel simulation) |
| R | A hundred to a thousand times | faster: putting both simulation and training on a single graphics card beats the conventional pipeline with CPU simulation and GPU neural networks by 2–3 orders of magnitude (technical report) |
Further reading
- Learning to Walk in Minutes (arXiv 2109.11978)
https://arxiv.org/abs/2109.11978 - Isaac Gym: high-performance GPU-based physics simulation (arXiv 2108.10470)
https://arxiv.org/abs/2108.10470 - MuJoCo documentation · Overview (what a physics engine is)
https://mujoco.readthedocs.io/en/stable/overview.html