NVIDIA Updates Isaac Lab 3.0 Beta 2, Stabilizing Newton Support and Multiple Physics Backends

NVIDIA’s Isaac Lab 3.0 Beta 2 stabilizes configurations for multiple physics backends, including PhysX, Newton and OVPhysX, as well as rendering and visualization. For reinforcement learning, it adds common training and playback commands that support rsl_rl, rl_games, skrl, sb3 and rlinf. The release notes say the software is in beta development for version 3.0 and that breaking changes and performance adjustments may be made before the final release.

For developers using multiple physics and rendering backends, Isaac Lab 3.0 Beta 2 is an update to coordinate environment selection, sensors and visualization while training. NVIDIA says the software is still in beta and may undergo breaking changes before the final release, so it should be treated as a stage for evaluating compatibility and workflows, not as a finalized platform for production use.

The focus of Beta 2 is not only on refining a single physics engine or rendering path, but also on improving the connections needed to select among multiple backends and run training and visualization. The release notes describe handling scene data and sensors across PhysX, Newton and OVPhysX, as well as validation and error messages for unsupported combinations of backends and renderers. Rather than adding features for a specific robot, the update appears to strengthen the foundation for training while switching simulation environments. However, because it is a beta, the presence of a feature should be distinguished from its status as part of a finalized stable release. In terms of how the components fit together, Newton presets for locomotion, manipulation and navigation, along with sensor support, are used to configure virtual environments. MCAP recording and playback and teleoperation handle operational data, while the common train/play commands and policy deployment guide are used to run training results. The release notes include these in the same 3.0 update, but do not say that data collection, training, deployment on physical robots and retraining have been automatically integrated into a single workflow. It is important to keep these boundaries clear. Another example of developer tools is a Physical AI News article on AWS releasing a toolchain for robot development. The AWS article covers the release of a developer toolchain, while the Isaac Lab release notes detail implementation updates to physics and rendering backends, training execution and teleoperation. Rather than grouping them together as announcements in the same product category, they should be compared by which layer of developers’ tools they address. Isaac Lab’s release notes also list supported configurations including Python 3.12, PyTorch 2.10.0 and CUDA, so the update covers installation environments as well as features. The next points to watch are how the differences in functionality among backends identified during the beta are addressed in the final release, and whether the release provides data on reproducibility and performance for the same task across the listed environments. The release notes provide no comparable performance figures or validation results on physical robots. The text also does not clarify which data formats or operating conditions will be standardized between MCAP recording and playback, training and policy deployment. Therefore, what can be assessed at this stage is an expansion of features and supported configurations, not a performance advantage or demonstrated applicability to physical robots.

Beta 2 expands Newton support for deformable bodies, coupling with rigid bodies, ray casters, IMU/PVA and contact sensors, among other features, and adds rough-terrain locomotion presets for quadrupedal and bipedal robots. It improves integration of scene data, sensors, cameras and visualization features across PhysX, Newton and OVPhysX. For rendering, it adds PPISP processing that converts HDR scene output to LDR camera images independently of the renderer backend. For teleoperation, the update supports MCAP recording and playback with Isaac Capture and revises workflows for imitation learning and policy deployment. Installation notes list common train/play commands, Python 3.12 and PyTorch 2.10.0. CUDA 12.8 is listed for x86_64/Windows, while CUDA 13.0 is listed for aarch64. The Isaac Sim Docker image listed in the release notes is pinned to version 6.0.0. A separate entry, v3.0.0-beta2.patch1, includes support for Isaac Sim 6.0.1 and notes fixes to dependency and Docker image compatibility.