MuJoCo 3.13.0 and Show-Harness | What makers can use

What actually changed in the last 24 hours? MuJoCo 3.13.0 is now a stable release with practical changes for simulating servos, springs and contact. Show-Harness is a preprint that explores teaching tasks across different robots through semantic actions and demonstrations collected with a browser and keyboard.

That distinction matters. MuJoCo can be installed today from PyPI. Show-Harness provides code, models and data, but its benchmarks come from the authors and have not been independently validated.

MuJoCo 3.13.0 makes larger timesteps easier to test

The main addition is the discrete integrator. According to the official release notes, it includes damping, position stiffness and constraint solving in the effective metric. The goal is to stabilize passive springs and position servos when using a larger simulation step, the amount of time advanced on each update.

The integrator documentation covers the mathematical context. A useful maker test is much smaller. Copy a scene that contains servos or springs, change only the integrator and compare stability, speed and contact behavior before modifying the timestep in the main project.

The official release notes identify one breaking change to check. Existing flex models that relied on the special implicit or implicitfast case with the CG solver now need integrator="discrete". If a deformable model behaves differently after the update, inspect that setting first.

The release notes also confirm the return of mjWARN_INERTIA and protection against singular mass matrices. Those safeguards were accidentally removed in MuJoCo 3.3.0. Their return helps expose models whose inertia leaves the simulated system poorly defined.

With elliptic cones and the Newton solver, the MuJoCo team reports an average speedup between 1.4 and 2 times. For slow steps in scenes with heavy sliding contact, it reports improvements between 3 and 4 times. These are the team’s own results, with no independent benchmark yet.

The official notes also list site meshes, multicontact cylinders, Python 3.15 support and a headless rendering sample that saves a PNG. The official MuJoCo 3.13.0 wheels are available, so the release can be installed and tested without building it locally.

Show-Harness turns semantic instructions into local motion

Show-Harness asks a vision-language policy to select discrete semantic actions. A robot-specific interpreter then converts those actions into local motion. This separates the high-level decision from the control needed by each platform.

Its GUMI interface records demonstrations through a browser and keyboard, without dedicated teleoperation hardware. The team evaluated the system on two physical platforms, a seven-degree-of-freedom Franka Research 3 and an AgileX dual-arm robot. There is no equivalent test on home hardware, so full reproduction remains outside a small build.

The published dataset contains 164 physical episodes with 7,774 steps. That total includes 101 Franka episodes and 63 AgileX episodes. It also provides 230 simulated episodes with 13,523 steps from ManiSkill and RoboLab.

For fine-tuning, the authors use Qwen3.5-2B with rank 64 LoRA. About 3% of the parameters are trained, with a reported duration under two hours on an H200. The paper also says the process fits on a 24 GB GPU, although it does not claim the same training time on that hardware.

The code, model weights and dataset are published under Apache-2.0. That makes the action interface and data format inspectable before committing to a complete robot setup.

The limits are clear. This is a preprint. Each task is tested ten times. Its zero-shot experiments depend on closed APIs, with Gemini 3.1 Pro as the main model. All reported results come from the authors and have no independent validation.

Where would I start? I would open the dataset and trace how one instruction becomes a discrete action. You do not need two industrial arms to understand that layer. Start with the format, one small task and an interpreter for a robot you already have.