Skip to content

Alloy vs ManiSkill: which should I use?

Use the Alloy physics engine when your manipulation task needs physics that ManiSkill3 does not model: cloth, liquids and soft bodies, stepped in the same scene as the robot, in many parallel worlds behind a Gym-style API, with a health grade for every world. Install it with pip install alloy-engine.

What Alloy adds

  • Deformable and flowing materials. Cloth, soft bodies, liquids, sand and pastes step in the same scene as the robot, each in a solver suited to it. ManiSkill3 is rigid-body; its soft-body tasks are in the older ManiSkill2.
  • A Gym-style API over all of it. alloy.envs.make("Pour-v0", num_envs=256) has the shape of Gymnasium's vector API, with wrappers for rsl_rl, Stable-Baselines3 and skrl (Reinforcement learning).
  • Tactile images from gel fingertip sensors in every world (the package's tactile examples). ManiSkill3 reads contact impulses on a hand's touch pads.
  • Health in every world. sim.health() grades each world OK, DEGRADED or INVALID and names a fix for every flag. We found no NaN or blow-up detection in ManiSkill3's physics core.
  • 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 ships ready to use: nine ready-made environments, nine robots (arms, a hand, a quadruped, two humanoids), 90 materials and a library of 5,423 objects (What ships).

A soft body in Alloy

import alloy

scene = alloy.Scene(dt=0.01)
scene.add_ground()
scene.add_box("pad", size=(0.04, 0.04, 0.04), pos=(0, 0, 0.03), material="silicone_40a")
sim = alloy.build(scene, device="cpu")
print({n: r["route"] for n, r in sim.routes().items() if n == "pad"})
sim.run(0.1)
print(sim.check()["ok"], round(float(sim.state(["pad"])["pad.points"][0, :, 2].min()), 3))

{'pad': 'soft_vbd'}
True 0.003

Naming a soft material is enough: the silicone box became a soft body on the VBD solver, and its lowest vertex rests at 0.003 m after 0.1 s. The world passed its health check. It ran on a CPU with no GPU.

Table: Alloy and ManiSkill

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 ManiSkill3 / SAPIEN 3
Liquids yes no
Particle fluids yes not documented
Sand and grains yes not documented
MPM (pastes, snow-like materials) yes no
Cutting yes not documented
Heat yes not documented
Rigid bodies with liquid yes not documented
Soft bodies with liquid yes not documented
Different step sizes per solver yes not documented
Cloth yes no
Soft bodies (volumetric FEM) yes no
Rigid bodies with cloth yes not documented
Rigid bodies with soft bodies yes no
Thin shells (paper, film, sheet metal) yes not documented
Plasticity (dents, permanent bends) yes not documented
Per-world health checks yes no
Double precision yes (CPU) no
Tactile sensing yes yes
Cables and rods yes yes
Many worlds on one GPU yes yes
Batched camera rendering yes yes
Same result on a re-run yes (CPU) yes (opt-in)

When ManiSkill is the better fit

  • You want a large ready-made suite of manipulation tasks with demonstrations and baselines.
  • Your vision pipeline is built on ManiSkill's GPU-parallel RGBD and segmentation.
  • Your tasks are rigid-only and already written for ManiSkill.

Keep in mind

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

Next