Which simulators handle deformable object manipulation: cloth, rope and soft bodies? Several open-source engines model some or all of them. They differ in which deformables they cover, whether those touch a rigid robot in the same scene, and what ships around the physics. Here is what each main open-source option is good for, as of late 2026.
What to check first
For a manipulation task, "supports cloth" is not enough. Check that the material you need (cloth, a cable or rope, a volumetric soft body) can touch a rigid robot in the same scene, that many worlds run at once on one GPU for reinforcement learning, and that you will know when a world blows up, so bad physics stays out of your training data.
The short table
The cells come from Alloy's compare pages, which read each engine's public code and docs on 2026-11-09: "yes" means the feature is in the public release and documented. They were read, not run.
| Alloy | Genesis | Newton | MuJoCo / MuJoCo Warp | MuJoCo Playground | ManiSkill3 | |
|---|---|---|---|---|---|---|
| Cloth | yes | yes | yes | yes | yes | no |
| Soft bodies (volumetric FEM) | yes | yes | yes | yes | yes | no |
| Cables and rods | yes | yes | yes | yes | yes | yes |
| Rigid bodies with cloth | yes | yes | yes | yes | yes | not documented |
| Rigid bodies with soft bodies | yes | yes | yes | yes | yes | no |
| Many worlds on one GPU | yes | yes | yes | MuJoCo Warp | yes | yes |
The table shows coverage, not quality: two "yes" cells can hide very different models, as the next section shows.
What each one is good for
Genesis (Apache-2.0) mixes materials in one scene and runs many worlds on one GPU. It builds a cable as a chain of jointed rigid links, and its user guide sets up cloth with a PBD material. Pick it when you already build on its API, renderer and tooling, or when you want double precision on a GPU: it has a 64-bit option.
Newton (Apache-2.0) provides solvers that cover cloth, soft bodies and cables, with many worlds on one GPU. Pick it when you want the solvers themselves or want to change their internals, or when you need differentiable simulation, inverse kinematics, or Newton's own sensors and viewers.
MuJoCo and MuJoCo Warp (Apache-2.0) document cloth, soft bodies and cables beside rigid robots. MuJoCo Warp runs many worlds on one GPU and is the leanest stack when the scene is rigid robots only. Work in MuJoCo itself when you want everything an MJCF file can set, or MuJoCo C alone on the CPU in double precision, with its viewer and tools.
MuJoCo Playground (Apache-2.0) offers a large set of ready-made locomotion, manipulation and dexterous-hand environments with tuned training recipes.
ManiSkill3 (Apache-2.0) is rigid-body; its soft-body tasks are in the older ManiSkill2. Its strength is a large ready-made suite of manipulation tasks with demonstrations and baselines, plus GPU-parallel RGBD and segmentation, and it fits best when your tasks are rigid-only and already written for it.
Alloy: cloth, cables and soft bodies beside a robot
Alloy (Apache-2.0), first released on PyPI in November 2026, is an open-source physics engine for robot learning. It simulates cloth, cables and soft bodies in one scene beside rigid bodies and robots, and sends each object to the solver that suits it:
- Cloth steps in Newton's VBD cloth solver, and folds and rests as two layers with self contact on.
- Soft bodies: give a box a soft material such as
silicone_40aor a foam, and the build can make it a soft body on Newton's VBD solver. - Cables are elastic rods that bend, twist and stretch as their material says. A rod that yields, such as a copper wire, keeps a bend. A clamped cable can also run on the fast chain of rigid segments.
- Robots step in MuJoCo Warp on a GPU and MuJoCo C on the CPU, so an MJCF robot loads directly and runs in the solver it was written for.
sim.routes() tells you which solver took each object. sim.check() and sim.health() grade every world OK, DEGRADED or INVALID and name a fix for every flag, so a physics failure is not read as a policy failure. Alloy also ships ClothFold-v0, a ready-made towel-folding environment, and documents cutting soft objects with a knife.
pip install alloy-engine This is step 1 of the docs page How do I simulate squeezing a soft object between two fingers?, copied as published:
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, 0.02), material="foam_pu_flexible") scene.add_box("left", size=(0.04, 0.01, 0.04), pos=(0, -0.026, 0.02), kinematic=True) scene.add_box("right", size=(0.04, 0.01, 0.04), pos=(0, 0.026, 0.02), kinematic=True) sim = alloy.build(scene) print(sim.routes()["pad"]["route"]) soft_vbd The material alone made the cube a soft body; the page then closes the fingers and reads its width back.
Where Alloy is weaker today: soft bodies step much slower than rigid ones, and cloth, liquid and soft bodies are much slower on a CPU. No friction grasp that carries a cotton sheet is verified yet. ClothFold-v0's opt-in grasp="attach" is a shortcut that holds the towel to the fingers, not a friction grasp; outside it, friction is the only grasp. ClothFold-v0 is experimental, slow and needs a GPU. It is a first release, so read its limits page before you trust a result.
How to choose
- Rigid robots only, and the leanest stack: MuJoCo Warp. Everything an MJCF file can set: MuJoCo.
- The solvers themselves, differentiable simulation or inverse kinematics: Newton.
- An existing Genesis codebase, or double precision on a GPU: Genesis.
- Ready-made rigid manipulation tasks with demonstrations: ManiSkill3.
- Cloth, cables and soft bodies beside an MJCF robot, one scene API, a health grade per world: Alloy.
Keep in mind
- In Alloy, sand and pastes do not build in one scene with fabric cloth, a soft body or a clamped cable. A soft body with a paper or film sheet, and fabric cloth with a paper or film sheet, are refused too. The build error names the nearest setup that does build (details).
- A cloth falling onto another cloth can pass through it, and a clamped cable beside a robot on the CPU steps slowly unless it uses the chain route.
- The other engines' cells were read, not tested: check the feature you need in their own docs.
- Full list of Alloy's limits: Keep in mind.
Links
- Docs page this is based on: https://alloy.argonrobotics.ai/docs/answers/deformable-object-manipulation/
- Engine-by-engine comparisons: https://alloy.argonrobotics.ai/docs/compare/
- Code: https://github.com/argonrobotics-ai/alloy
- Paper: https://arxiv.org/abs/2611.16312

