Physical AI Lab // Episode 1.4 // One-pager

A map of simulators

MuJoCo, Isaac, Newton, Genesis: what each does best

Three key points

  1. At its core, a simulator computes the forces at the moment of contact, step by step; engines approximate contact differently, so the same motion's trajectories diverge.
  2. Four engines, four trade-offs: MuJoCo computes contact accurately (DeepMind, open source); Isaac is fast through GPU parallelism (the technical report says 2–3 orders of magnitude faster than conventional CPU simulation with GPU network training, R); Newton is new and extensible (NVIDIA / Google / Disney announced joint development; NVIDIA says MuJoCo-Warp makes humanoid simulation more than 70× faster, O); Genesis aims to unify many kinds of physics (very new: check how well it has been validated).
  3. Speed lets reinforcement learning do huge amounts of trial and error (the paper reports thousands of parallel environments on a single GPU; the ANYmal four-legged robot learns to walk on flat ground in under 4 minutes, R). But fast is not the same as accurate, and simulation is not the real world (the sim-to-real gap): choose an engine by the task.

Evidence: O official · R research · M media

R2–3 orders of magnitudefaster, for physics simulation and training together on a single GPU, than conventional training with CPU simulation and GPU networks (NVIDIA Isaac Gym technical report)
O70×faster: Google DeepMind's MuJoCo-Warp speeds up humanoid simulation by more than 70×, and in-hand manipulation by about 100× (NVIDIA says)
Runder 4 minutesof training for the ANYmal four-legged robot to learn to walk on flat ground, with thousands of parallel environments on a single workstation GPU; about 20 minutes on uneven terrain (paper's own report)

Further reading