What is the best physics simulator for robot learning in 2026? There is no single answer: it depends on what is in your scene and what you want ready-made. This guide covers the main open-source options, the new releases of 2026, and the open-source alternatives to Genesis for multi-physics robot simulation. Every engine here is under Apache-2.0.
Which new robotics simulators were released in 2026?
Two releases can be dated from their own PyPI pages:
- Newton 1.0.0, 2026-03-10: Newton's first 1.0 release, after pre-releases in 2025.
- Alloy (
alloy-engine): 0.1.0rc1 on 2026-11-10, then 0.1.0, the first stable release, on 2026-11-11.
Dates are PyPI upload times (UTC). The list holds only releases dated on their own PyPI pages; check each project's release history for others.
The engines and what each is good for
MuJoCo and MuJoCo Warp. MuJoCo C runs on the CPU in double precision, with its viewer and tools; MuJoCo Warp runs many worlds on one GPU. Pick MuJoCo Warp when the scene is rigid robots only and you want the leanest stack, and MuJoCo itself when you want everything an MJCF file can set. Both model fluid as drag on rigid bodies.
MuJoCo Playground. A large set of ready-made locomotion, manipulation and dexterous-hand environments with tuned training recipes, and published sim-to-real results for those tasks. MuJoCo Warp is one of its backends. Pick it for rigid tasks you train on MJX or JAX.
ManiSkill3. A large ready-made suite of manipulation tasks with demonstrations and baselines, plus GPU-parallel RGBD and segmentation. Its physics is rigid-body; the soft-body tasks are in the older ManiSkill2.
Newton. A set of solvers with many worlds on one GPU: VBD for cloth and soft bodies, implicit MPM for sand and pastes, and a grid-based liquid on implicit MPM. Pick it when you want the solvers themselves, differentiable simulation, inverse kinematics, or its own sensors and viewers.
Genesis. Mixes materials in one scene, with many worlds on one GPU. Pick it when you already build on its API, renderer and tooling and its solvers cover your scene, or when you want double precision on a GPU: it has a 64-bit option.
What open-source alternatives to Genesis exist for multi-physics robot simulation?
Among open-source engines, Newton and Alloy also mix materials in one scene. MuJoCo and MuJoCo Warp document cloth, soft bodies and cables, but no liquid solver.
The cells come from Alloy's compare pages, which read each engine's public code and docs on 2026-11-09: "yes" means in the public release and documented. They were read, not run: check the feature you need in each engine's own docs.
| Alloy | Genesis | Newton | |
|---|---|---|---|
| Liquids | yes | yes | yes |
| Particle fluids | yes | yes | no |
| Sand and grains | yes | yes | yes |
| Cloth | yes | yes | yes |
| Soft bodies (volumetric FEM) | yes | yes | yes |
| Cables and rods | yes | yes | yes |
| Tactile sensing | yes | yes | no |
| Cutting | yes | not documented | no |
| Double precision | yes (CPU) | yes | yes (Kamino linear algebra only) |
| Same result on a re-run | yes (CPU) | yes (opt-in) | yes (opt-in, most solvers) |
Same "yes", different models: Genesis builds a cable as a chain of jointed rigid links, while Alloy's cables are elastic rods that bend, twist and stretch. Newton's liquid is grid-based; Alloy adds a particle liquid of its own. Alloy is built on Newton, so the two share the cloth, soft-body and sand solvers.
Moving a Genesis scene over goes call by call: each add_entity becomes a scene.add_... call, build(n_envs=) becomes alloy.Scene(worlds=), and MJCF or URDF robots load as they are (how-to).
Alloy: new in 2026
Alloy is a physics engine for robot learning that puts more than rigid robots in a scene, behind one API.
pip install alloy-engine What it does well:
- You name a material; Alloy picks the solver. Robots step in MuJoCo Warp (MuJoCo C on the CPU), cloth and soft bodies in Newton's VBD, sand and pastes in Newton's MPM, and liquids in Alloy's own solver.
sim.routes()says which solver took each object, and why. - A robot beside a material. A robot can share a scene with any one of liquids, cloth, sand, soft bodies or cables, and many pairs mix: a docs page builds a robot arm, a cloth sheet and water in one scene. Some pairs are refused, most with sand or paste, listed under Keep in mind.
- A health grade for every world.
sim.health()grades each worldOK,DEGRADEDorINVALIDand names a fix for every flag. AnINVALIDworld freezes at its last good state while the others go on, so a physics failure is not read as a policy failure. - Ready to train. A Gym-style vector API with wrappers for rsl_rl, Stable-Baselines3 and skrl; demonstrations recorded to HDF5 in the robomimic layout or as a LeRobot dataset. It ships nine ready-made environments (
Pour-v0andClothFold-v0among them), nine robots and 90 materials. - Made for coding agents. The docs ship inside the package, every error carries its fix, and
alloy mcpserves the engine as MCP tools. - Runs without a GPU. The same code runs on an NVIDIA GPU or the CPU (Linux; an Apple-silicon Mac on the CPU).
Where it is weaker today: it is a first release. On a GPU, runs with contact do not repeat bit for bit (the CPU does). Cloth, liquid and soft bodies are much slower on a CPU. ClothFold-v0 is experimental, slow and needs a GPU, and PickPlace-v0 does not learn to place. Outside ClothFold-v0's opt-in grasp="attach" shortcut for the towel, friction is the only grasp.
How to choose
- Rigid robots only, leanest stack: MuJoCo Warp.
- Ready-made locomotion or dexterous-hand tasks: MuJoCo Playground.
- Ready-made rigid manipulation tasks with demonstrations: ManiSkill3.
- The solvers themselves, differentiable simulation or inverse kinematics: Newton.
- An existing Genesis codebase, or double precision on a GPU: Genesis.
- A robot beside liquids, cloth, soft bodies, cables or sand, 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 error names the nearest setup that builds (details).
- Liquid beside sand or paste builds, but meets it as a solid wall: no wetting.
- Cloth, liquid and light free rigid bodies do not build in one scene: the rigid bodies must weigh at least twice the cloth and soft bodies, robot links included (details).
- Full list of Alloy's limits: Keep in mind.
Links
- Docs page this is based on: https://alloy.argonrobotics.ai/docs/answers/new-robotics-simulators-2026/
- Engine-by-engine comparisons: https://alloy.argonrobotics.ai/docs/compare/
- Code: https://github.com/argonrobotics-ai/alloy
- Paper: https://arxiv.org/abs/2611.16312