What happened
Google DeepMind’s MuJoCo Robotics Simulation Team announced MuJoCo 4.0 on November 7, 2026, together with MuJoCo Playground 1.0. The announcement says the project receives about 3.6 million PyPI downloads per month, has more than 16,000 GitHub stars, and drew around 2,000 commits from more than 45 external contributors over the previous year. MuJoCo 4.0 is the major release after 3.0 in 2023; since 3.5, the middle version number marks breaking changes.
The most visible developer-facing change is MuJoCo Studio. Its composable SDK for Python and C++ combines simulation, Dear ImGui interfaces, and 3D rendering without requiring teams to extend the older fixed-function OpenGL viewer, which is now in maintenance mode. Studio separates physics from UI rendering through non-blocking message channels in Python, and supports hot-reloading viewer plugins, simulation plugins, and MJCF files. The optional Studio App can run natively or through a browser session at live.mujoco.org. A remote or GPU-less machine can host a session; one browser user at a time controls the simulation and interface, while others watch and control can be handed over. The browser viewer accepts XML, MJB, and MJZ models, including deep links to models in GitHub repositories.
MuJoCo 4.0 also moves soft bodies from experimental plugins and explicit penalties to native continuum deformables solved implicitly alongside rigid bodies, joints, and tendons. The announcement describes cloth, shells, 3D solids, and elastic rods, plus plasticity for joints, tendons, shells, and solids. It says the implicit formulation can keep stiff servos, cables, thin shells, and 3D solids stable at timesteps 5–10 times larger. Penetration-free cloth contact and plastic deformation are available through model settings and elements. Large-strain Neo-Hookean solids remain experimental.
The collision system adds speculative contacts, which sweep fast-moving geometry ahead to prevent tunneling without sub-stepping. Google DeepMind reports that a robot arm caught a 0.7 mm playing card falling at over 2 m/s with a 16 ms timestep; discrete collision detection missed it above 8 ms. Humanoid stair climbing and thin-object grasping ran 6–9 times faster in the team’s tests. The feature currently covers convex primitives and meshes on CPU; heightfields, signed distance fields, flexes, MJX, and MuJoCo Warp are listed as future work. Packaging also changes: mujoco-core provides a minimal headless install, while mujoco installs the optional components too. Python 3.10 is no longer supported, and upgrading from 3.x requires uninstalling the old package first.
What the archive adds
New Runtime covered Hi3D 1.0, an open-source image-conditioned 3D mesh refinement release on October 24, 2026, and AnyMove, an open-source creature-motion capture pipeline on November 7, 2026. Those are adjacent 3D and simulation-facing tools, not earlier coverage of MuJoCo; this is a new event and a distinct robotics simulation thread.
Trend signal
This release fits agent-ready software exposes capabilities only weakly: Studio’s composable SDK and plugin interfaces make simulation and visualization capabilities easier to extend, but the announcement does not establish an agent-specific interface or permission model. The stronger signal is a more modular developer surface, not evidence that the simulator is already designed for external agents.
Why now
The team says it saved version 4.0 for a major step after 3.5 made the middle version number the marker for breaking changes. The announcement frames the release as the culmination of changes across 39 releases since MuJoCo 3.0, including a rebuilt collision stack and increasingly mature deformable simulation.
Why it matters
For robotics teams, simulation quality and iteration speed are linked: a model that tunnels through a thin object or needs very small timesteps can make testing slower or physically misleading. Speculative contacts target that failure mode, while the reported card-catching example gives builders a concrete scenario to reproduce rather than a general claim about accuracy. The larger-timestep results could also reduce the computational cost of some simulations, though the announcement does not quantify total runtime or resource savings across workloads.
Studio addresses a separate bottleneck: building a useful visualization or debugging interface has historically meant extending a monolithic viewer or maintaining custom tooling. A composable SDK, hot reload, and browser sharing could make it easier to inspect models and collaborate on remote experiments. The split package gives teams a smaller headless option, but upgrading has real migration implications: Python 3.10 is dropped, old packages must be removed first, and soft-body APIs and viewer paths have changed. Teams should test representative models and workflows before treating the new capabilities as drop-in replacements.
What happens next
- More MuJoCo simulation features will become available across heightfields, SDFs, flexes, MJX, and MuJoCo Warp. (6 months, 75%). Watch for: release notes documenting speculative-contact support beyond CPU convex primitives and meshes.
Limits and uncertainty
The performance and robustness figures are results reported by Google DeepMind; the announcement does not provide an independently run evaluation or broad workload methodology. It does not quantify the compute cost of the new deformable solver or the overall speedup teams should expect in their own environments. Speculative contacts have explicitly limited coverage at launch, with several geometry types and accelerator paths described as future work. The announcement lists migration-sensitive API changes, but teams will need to assess compatibility against their own models and plugins.
Skill to build now
Take one scene in which your robot interacts with a thin or fast-moving object and compare it under discrete collision detection and speculative contacts. Record the timestep, whether tunneling occurs, contact stability, and runtime; then test a larger timestep while checking that the behavior still matches your task’s acceptance criteria. This isolates whether the new mode helps your workload rather than assuming the release’s example transfers directly.
Reader takeaway
MuJoCo 4.0 is more than a new viewer: it combines a modular visualization surface with meaningful changes to collision handling and deformable simulation. The reported gains are promising but workload-dependent, so the practical next step is to test the new contacts and migration path on representative scenes before relying on them in production.