Genesis 1.4.1 and wbc-mjlab 0.0.6 | Robotics

What actually changed in the last 24 hours? Two software releases are worth checking if you work with simulation or humanoid control. Genesis 1.4.1 targets large-scene performance and fixes import and actuator behavior. wbc-mjlab 0.0.6 introduces an experimental Unitree G1 control path that does not retarget joints at runtime.

The maturity level is different. Genesis 1.4.1 is an installable simulator update. The wbc-mjlab path is research software, not a product or independent evidence of a policy working on a physical robot.

Genesis 1.4.1 focuses on large scenes and fixes simulation-breaking details

Genesis 1.4.1 was published on September 12 at 18:58:55 UTC. Its PyPI wheel followed at 18:59:45 UTC and supports installation with pip on Python 3.10 through 3.13.

The maintainer reports better performance for large scenes on both CPU and GPU. The official notes name three cases: simulations split into independent groups, Jacobi solving and contact calculations on the CPU. The release includes the project’s own comparison chart, but no independent benchmark. Treat it as a speedup measured by Genesis in its environment, not one that transfers automatically to every scene.

Several smaller fixes can change visible results. Normalized glTF accessors store integers that software must convert to a decimal range when reading them. Genesis now performs that conversion correctly and preserves the normals and UV coordinates that define mesh orientation and texture. The release also honors MJCF actuator force ranges and lets exported scenes open on any platform through gs launch, according to the official changelog.

The API now provides a public method for setting degree-of-freedom limits on each entity. This avoids dropping down to the solver and using scene-wide indices to change those limits, as documented in the release notes.

My practical test would be small. Create an isolated environment, pin genesis-world==1.4.1 and run a copy of a scene with several bodies and contacts. Check step time, actuator force limits and glTF mesh appearance. That covers the actual changes without risking a working project.

wbc-mjlab 0.0.6 removes runtime joint retargeting from its new G1 path

wbc-mjlab 0.0.6 was published on September 13 at 07:50:36 UTC. Its main addition is a first experimental deployment path for the Unitree G1 without joint retargeting at runtime. In plain terms, the actor, the part of the policy that chooses the next action, does not consume a newly retargeted joint trajectory while it runs.

Instead, the actor receives relative 3D positions and 6D orientations for the wrists and ankles. Training still uses whole-body control, or WBC, to coordinate all the joints. The critic, the component that evaluates actions during training, sees the complete references, according to the official task guide and release notes.

The package adds Wbc-G1-EE and Wbc-G1-EE-SE, sample commands for training and playback, and YAML configuration export. It requires mjlab>=1.6.0, as specified in the official guide.

The limit matters more than the headline. This targets a Unitree G1, not home workshop hardware. The guide documents the training and deployment flow, but neither it nor the release provides independent physical validation. Treat it as experimental software. It is not evidence by itself that the policy works on a real G1.

You can still learn from it without owning the humanoid. Start with the YAML export and the end-effector reference format. Then replay one task in simulation and inspect what the actor observes versus the extra information available to the critic. That gives you a useful, limited test before you consider expensive hardware.