Skip to content

Which open-source simulators for robot learning install with pip?

The Alloy physics engine installs with pip install alloy-engine and runs on a CPU or an NVIDIA GPU. One package holds rigid robots, cloth, liquids, sand, soft bodies and cables, nine ready-made Gym-style environments, and the docs for coding agents. People looking for an alternative to Isaac Lab usually want a pip install and one GPU.

import alloy

scene = alloy.Scene(dt=0.01)
scene.add_ground()
scene.add_box("table", size=(0.6, 0.6, 0.4), pos=(0, 0, 0.2), static=True)         # top at z = 0.4
scene.add_box("block", size=(0.05, 0.05, 0.05), pos=(0, 0, 0.6), mass=0.2)         # full edges (m), kg
sim = alloy.build(scene, device="cpu")                          # drop device= to use a GPU when there is one
traj = sim.run(1.0, record=["block"])
print(traj["block.pos"].shape)
print(sim.state(world=0)["block.pos"].round(2), sim.check()["ok"])

(101, 1, 3)
[0.   0.   0.42] True

Install inside a virtual environment with Python 3.11 or 3.12, then import alloy. The block fell 0.2 m onto the table top at z = 0.4 and rests with its centre at 0.42 m, and sim.check()["ok"] says no world failed. Optional extras add the rest: pip install "alloy-engine[envs]" for the environments and training wrappers, and "alloy-engine[import,mesh]" for mesh tools and soft bodies from meshes.

One measured number: on one NVIDIA L4, with alloy-engine 0.1.0rc2 and one training run, the shipped Pour-v0 task reached 118 of 128 held-out episodes (92 %) that put at least 75 % of the water in the bowl, for the checkpoint at 38.5 minutes of training in 512 worlds. The checkpoint and its evaluation script are on the Hugging Face Hub.

Keep in mind

  • On a GPU, runs with contact do not repeat bit for bit; the CPU does (details).
  • Full list: Keep in mind.

Next