🤖 Google DeepMind has released MuJoCo 4.0, the first major version of the physics engine for robotics in three years.
MuJoCo is a simulator used to train and test robots before deploying them in the real world. It's used by almost everyone involved in robotics and RL: about 3.6 million downloads from PyPI per month and over 16,000 stars on GitHub.
What's changed compared to 3.0:
- Fast and thin objects: previously, the simulation step had to be reduced to prevent them from passing through each other, now speculative contacts solve this
- Soft bodies: instead of experimental support with a small step - cloth without penetrations, ropes that bend and twist, and plasticity (wire, clay, and foil retain dents)
- Large scenes: groups of objects that do not touch each other are calculated separately by default, stationary objects "fall asleep," and calculations are done in multiple threads
- Where it works: previously CPU plus GPU/TPU via MJX, now also MuJoCo Warp with NVIDIA and directly in the browser
- Visualization: instead of the old viewer on OpenGL 1.5 - MuJoCo Studio with realistic rendering
- Robot models: about 70 robots in the Menagerie library instead of 20
The main feature is speculative contacts. The simulator calculates physics in steps, and if an object travels more than its thickness in one step, it passes through an obstacle. Now the engine "predicts" the trajectory of fast objects in advance and catches collisions. In team tests, a hand catches a playing card 0.7 mm thick, falling at a speed of over 2 m/s, with a step of 16 ms. The old approach missed already at a step greater than 8 ms.
Why this matters. Robots are trained in simulation for billions of steps, and the speed of simulation directly determines the cost of training. Climbing stairs with a humanoid and grasping thin objects are now calculated 6-9 times faster because a larger step can be taken. And cloth, ropes, and soft materials are exactly what robots will have to work with at home: laundry, wires, packaging. The figures are currently measured by the team itself, but the direction is clear: the more accurate and cheaper the simulation, the faster robots come to the real world.
https://github.com/google-deepmind/mujoco/discussions/3773
https://x.com/_bbelousov/status/2119200725008024035
#robotics@rvnikita_blog #deepmind@rvnikita_blog #google@rvnikita_blog #simulation@rvnikita_blog #ai@rvnikita_blog