Alloy: A Multi-Material Physics Engine for Robot Learning
Organizations: Argon Robotics
Abstract
Robots handle much more than rigid objects: towels, cables, water, sand. The simulators used to train robots at scale are strongest on rigid bodies, and adding another material usually means a second engine and hand-written glue between the two. We present Alloy, an open-source physics engine for robot learning in which rigid robots, cloth, cables, liquids, granular material and soft bodies share one scene. Each object is described once; Alloy assigns it to a solver suited to its material, couples the solvers inside every control step, and runs the scene in many parallel worlds on one GPU, or on a CPU. Every world carries a health grade, to help tell a failure of the physics from a failure of the policy. Alloy is built on Newton, MuJoCo Warp and Warp, and it is written to be driven by coding agents: its documentation ships inside the package, every error names its fix, and the engine is exposed as tools over the Model Context Protocol. In a cold test, agents that had never seen Alloy installed a pre-release from the package index and trained policies for walking, cable handling and pouring on a single GPU. We also report how throughput grows with the number of worlds, depth-camera throughput, agreement with closed-form references, and a reliability test over a thousand worlds.
Figures & tables
| Solver | Objects routed to it |
|---|---|
| MuJoCo Warp (GPU), MuJoCo (CPU) | rigid bodies, robots, unclamped chains (such as cables), creased panels |
| VBD (Newton) | cloth, soft volumes, elastic rods |
| Style3D (Newton) | anisotropic cloth when it is the only dynamic object |
| Film (Alloy, on VBD) | paper, polymer film and metal sheets |
| Implicit MPM (Newton) | sand, pastes, viscous or yielding fluids |
| SPH (Alloy) | liquids |
| Task | Training | Training time | Held-out success |
|---|---|---|---|
| Walking : Unitree Go2, shipped environment, flat ground | PPO, shipped trainer, 4,096 worlds | 4.1 min 15.5 min | 244/256, 95 % [92, 97] 497/512, 97 % [95, 98] |
| Push recovery : the same policy, pushed with 150 N for 0.1 s | as above | 15.5 min | 252/256 and 248/256, 97–98 % |
| Cable lift : task written by the tester | PPO, shipped trainer, 1,024 worlds | about 15 min about 30 min | 242/256, 94.5 % [91, 97] 245/256, 95.7 % [93, 98] |
| Pouring : shipped environment | PPO, 512 worlds | 38.5 min | 118/128, 92 % [86, 96] |
| Cup pouring : task written by the agent, on its second reward version | PPO, 256 worlds | about 70 min | 256/256, 100 % [98.5, 100] |
| Pick-and-place by imitation : shipped environment, gantry gripper, state observations | behaviour cloning, 1,000 scripted demonstrations | 66 s | 256/256, 100 % [98.5, 100] |
| Depth images per second, run 1 | 70,786 | 16,111 | 4,133 |
| Depth images per second, run 2 (another machine) | 62,785 | 14,372 | 4,106 |
| Check | Reference | Error | Tolerance |
|---|---|---|---|
| Pendulum period † | small-angle period | ||
| Momentum | conservation of momentum | ||
| Sliding with friction | Coulomb friction law | ||
| Rope around a post † | capstan equation | exact | |
| Rod bending under a tip load | cantilever deflection | ||
| Rod twisting | torsion formula |
Robotics Simulation
Robotics simulation builds virtual robots and environments to model, train, and evaluate autonomous systems before real-world deployment.
See all 339 papers