Discussion

MuJoCo 3.14 adds penetration-free flex contact, but changes a key actuator API

In Developer Tools

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#3270

Google DeepMind’s MuJoCo 3.14.0 release adds an experimental contact mode designed to stop flexible simulated objects tunnelling through one another, while also introducing a breaking actuator-control API change. That makes this useful robotics infrastructure news, not merely another version number wearing a tiny hat.

Google DeepMind Watch analysis

What happened

Released on 22 September, MuJoCo 3.14.0 introduces the ipc contact mode for supported two-dimensional flexible bodies. The release says each step verifies committed position updates with continuous collision detection, so flex contact cannot tunnel. The mode is frictionless and does not support exact replay through the normal mjgetState and mjsetState state capture, because it retains contact multipliers that those functions do not record.

The release also adds archive resource providers for containers such as .mjz and .zip, preserves frame transforms when saving MJCF models, and makes procedural model editing substantially faster. Building a specification with 8,000 geoms is reported as roughly 10 times faster, according to the release notes.

The sharp-edged item is a breaking API change to mjreadCtrl. It now follows the signature and semantics of mjreadSensor, including output buffers sized for multi-input actuators. Actuator history buffers also store multiple control values per sample, and mjinitCtrlHistory now expects the corresponding larger value set.

Why it matters

Reliable contact handling is important for robotics simulation because a flexible object passing through another object can produce physically implausible training data and misleading controller behaviour. MuJoCo’s new mode targets that problem directly, although its experimental status, frictionless contacts and incomplete replay state mean it is not a universal drop-in replacement.

The API change matters to developers maintaining control, actuator-history or multi-input-actuator code. Existing integrations should be checked and tested rather than upgraded on autopilot. A simulator that becomes more physically convincing while quietly changing the plumbing is exactly the sort of gift that arrives with a migration guide-shaped bow.

Our read

This is a substantial engineering release for robotics teams, particularly those modelling flexible bodies or building large procedural scenes. The practical recommendation is to test ipc on representative contact-heavy models, while treating the new actuator API as a migration task rather than a routine patch.

The release’s biggest promise is not that every simulation becomes magically lifelike. It is that two long-standing sources of friction, contact tunnelling and inefficient model editing, get serious attention. The useful next step is measurement: compare stability, runtime and controller behaviour on your own workloads before declaring victory.

What to watch

  • Whether the experimental contact mode gains friction support and broader flex-model coverage.
  • How teams adapt code using mjreadCtrl, actuator history buffers and multi-input actuators.
  • Whether the reported speed-up appears on larger, mixed model-building workloads.
  • Whether the new replay-state limitation changes how researchers reproduce experiments.

Discussion spark: Should robotics teams adopt an experimental simulator feature that prevents physically impossible contact, even when it changes replay semantics, or should reproducibility come first?

Sources and evidence
  • 3.14.0 (22 September 2026, 16:52 UTC)

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.