Skip to content

Alloy vs MuJoCo Playground: which should I use?

Use the Alloy physics engine when the task has more than rigid bodies: pouring a liquid or moving sand, or starting from a ready-made cloth-folding task (ClothFold-v0), in many parallel worlds with a camera in each. It steps robots in MuJoCo Warp, one of Playground's backends, and installs with pip install alloy-engine.

What Alloy adds

  • Tasks with liquid and cloth. Pour-v0 and ClothFold-v0 are two of the nine ready-made environments, and alloy.envs.make builds each in as many worlds as you ask for. Playground's engine models fluid as drag on bodies; we found no liquid, sand or paste solver in it.
  • The same robot solver. Alloy steps rigid bodies and robots in MuJoCo Warp, and an MJCF robot loads directly.
  • Health in every world. sim.health() grades each world OK, DEGRADED or INVALID and names a fix for every flag, so a physics failure is not counted as a policy failure. Playground has overflow and bad-acceleration warnings.
  • Cutting of soft objects, heat in liquids, and a different step size for each solver in one scene.
  • It is made for coding agents: the docs ship inside the package (python -m alloy.docs), every error carries its fix, and alloy mcp serves the engine as MCP tools (agents). It also records demonstrations (Imitation learning).

A pour in two worlds

import alloy

scene = alloy.Scene(worlds=2, dt=0.01)
scene.add_ground()
scene.add_open_box("cup", inner=(0.05, 0.05, 0.05), wall=0.01, pos=(0, 0, 0))
scene.add_pour("stream", 2e-5, pos=(0, 0, 0.1))                 # 20 ml, straight down from 10 cm
sim = alloy.build(scene, device="cpu")
sim.run(1.0)
points = sim.state(["stream"])["stream.points"]
in_cup = (abs(points[..., 0]) < 0.025) & (abs(points[..., 1]) < 0.025) & (points[..., 2] < 0.06)
print(len(points), in_cup.mean(axis=1).round(2))
print(sim.check()["ok"], [w["severity"] for w in sim.health()["worlds"]])

2 [0.99 0.99]
True ['OK', 'OK']

Two worlds each poured 20 ml of water into a cup, stepped together; the first line is the world count and the share of the water inside the cup in each world, and each world got its own health grade. Pour-v0 wraps a pour like this as a ready-made environment (pip install "alloy-engine[envs]"). It ran on a CPU with no GPU.

Table: Alloy and MuJoCo Playground

One rule for every engine: "yes" means the feature is in the engine's public release and documented; "not documented" means our survey has no public source for it. A note in brackets says where a "yes" holds. The other engines' cells were read from their public code and docs on 2026-11-09, not run by us. The scorecard file, bench/scorecard/row6/capabilities.jsonl in Alloy's source repository, has 26 rows and also grades how complete each feature is; this table shows the 22 rows Alloy's public package documents.

Capability Alloy MuJoCo Playground
Liquids yes no
Particle fluids yes no
Sand and grains yes no
MPM (pastes, snow-like materials) yes no
Cutting yes no
Heat yes not documented
Soft bodies with liquid yes no
Different step sizes per solver yes no
Same result on a re-run yes (CPU) not documented
Rigid bodies with liquid yes yes
Cloth yes yes
Soft bodies (volumetric FEM) yes yes
Rigid bodies with cloth yes yes
Rigid bodies with soft bodies yes yes
Tactile sensing yes yes
Thin shells (paper, film, sheet metal) yes yes
Plasticity (dents, permanent bends) yes yes
Cables and rods yes yes
Per-world health checks yes yes
Many worlds on one GPU yes yes
Batched camera rendering yes yes
Double precision yes (CPU) yes (jax_enable_x64)

When MuJoCo Playground is the better fit

  • You want a large set of ready-made locomotion, manipulation and dexterous-hand environments with tuned training recipes.
  • You want to start from published sim-to-real results for those tasks.
  • Your task is rigid-only and you want the fastest training on MJX or JAX.

Keep in mind

Alloy's limits are listed on one page, each with its size and a workaround: Keep in mind.

Next