Many of the tasks we want robots to learn mix materials: pouring water into a glass, folding a towel, scooping sand, routing a cable. In simulation those materials usually live in different tools, so a policy that has to pour and grasp ends up stitched across engines, or the liquid gets left out.

Alloy is an open-source (Apache-2.0) physics engine for robot learning that puts them in one scene. The paper is on arXiv: Alloy: A Multi-Material Physics Engine for Robot Learning.

A robot arm pours water into a glass

One scene, one solver per material

Alloy is built on MuJoCo Warp and Newton and picks a solver for each object from its material:

Object Solver
Robots and rigid bodies MuJoCo Warp (MuJoCo C on the CPU)
Cloth and soft bodies Newton's VBD solver
Sand, granular material and pastes Newton's implicit MPM solver
Liquids Alloy's own DFSPH solver

Objects from different solvers touch each other in the same scene. sim.routes() tells you which solver each object got. Here an MJCF arm sits beside a tub of water and a tray of sand:

import alloy ARM = """<mujoco><compiler angle="radian"/><worldbody><body name="link1" pos="0 0 0.1"> <joint name="shoulder" type="hinge" axis="0 1 0" range="-2 2" damping="2"/> <geom type="capsule" fromto="0 0 0 0.3 0 0" size="0.02" mass="1"/> </body></worldbody> <actuator><position joint="shoulder" kp="200" kv="10"/></actuator></mujoco>""" scene = alloy.Scene(dt=0.01) scene.add_ground() scene.add_robot("arm", mjcf=ARM, pos=(0, 0, 0.1)) scene.add_open_box("tub", inner=(0.08, 0.08, 0.05), wall=0.01, pos=(0.5, 0, 0)) scene.add_liquid("water", size=(0.064, 0.064, 0.024), pos=(0.5, 0, 0.024)) scene.add_open_box("tray", inner=(0.1, 0.1, 0.05), wall=0.01, pos=(0.5, 0.4, 0)) scene.add_box("sand", size=(0.06, 0.06, 0.03), pos=(0.5, 0.4, 0.03), material="sand_silica_dry") sim = alloy.build(scene, device="cpu") print({n: sim.routes()[n]["solver"] for n in ("arm", "water", "sand")}) sim.run(0.3, targets={"shoulder": (0.0, 0.3)}) print(sim.state(world=0)["arm/shoulder.q"].round(2), sim.check()["ok"]) 
{'arm': 'mujoco', 'water': 'sph', 'sand': 'mpm'} 0.25 True 

The MJCF string is the robot; a robot file such as a Franka loads the same way with add_robot. sim.check() grades the world, so you can catch a simulation that blew up instead of training on it.

A towel being folded

For reinforcement and imitation learning

  • Many copies of a scene run in parallel worlds on one NVIDIA GPU, or on a CPU.
  • Gymnasium-style vector environments and ready-made training tasks.
  • Trained example policies on the Hub: pouring, cable lifting and Go2 walking.

Four robot arms place boxes in bins

For coding agents

Alloy ships an MCP server, so an agent such as Claude Code can build and run scenes itself:

pip install alloy-engine claude mcp add alloy -- alloy mcp 

When to use something else

  • Rigid robots only: MuJoCo Warp on its own is the leaner choice.
  • If you want the solvers themselves, differentiable simulation or inverse kinematics, use Newton directly.
  • Many worlds of sand or paste are limited by the stress solve, and a liquid beside sand meets it only at the sand's surface. The full list is on the Keep in mind page.

Links