As of August 7, 2026, the official Isaac Lab release is v3.0.0-beta2.patch1 with Isaac Sim 6.0.1 support, while the Newton project releases have advanced from the 1.0 general-availability baseline to v1.4.0. Treating both as a single stable 'version 1' installation would mix two release tracks and can produce stale commands or incompatible dependencies.
Choose PhysX when you need the broadest current Isaac Sim feature coverage; test Newton's MuJoCo-Warp backend when kit-less execution, MuJoCo semantics or GPU-parallel reinforcement learning is the priority and your environment has a Newton configuration. The simulator comparison guide provides the wider landscape.
Start by naming the two release tracks correctly
The current v3.0.0-beta2.patch1 release remains on the Isaac Lab 3.0 beta line and adds Isaac Sim 6.0.1 support. The release notes warn that development remains active and that behavior, errors or performance can change before a stable release.
Newton 1.0 reached general availability in March 2026, but it is not the latest Newton package on the check date. The official repository lists v1.4.0, released July 16, with updated Warp and MuJoCo-Warp dependencies, so pinning an old 1.0 command as 'current' is unsafe.
| Track | Status on Aug. 7, 2026 | Planning consequence |
|---|---|---|
| Isaac Lab | v3.0.0-beta2.patch1 | Expect active changes; pin the beta and read release notes |
| Isaac Sim | beta2.patch1 supports Isaac Sim 6.0.1 | Match the supported application/runtime |
| Newton | v1.4.0 current official release | 1.0 is the GA milestone, not the current package |
| MuJoCo-Warp dependency | Version-bounded by Newton 1.4.0 | Use the release's dependency range rather than guessing |
PhysX remains the default; verify support by task
The `isaaclab_physx` backend is Isaac Lab's default PhysX path. Beta 2 expanded Newton support for deformables, ray casting, IMU/PVA sensing, contacts and joint torque, so these capabilities should not be labeled broadly as PhysX-only; verify each task preset, sensor, renderer and backend combination.
For an existing Isaac Sim scene, PhysX can be the first baseline. If Newton is being evaluated, use a preset explicitly supported by the current release and record unsupported combinations. Neither backend automatically makes contacts, materials, actuators or timing match a real robot.
Newton brings MuJoCo-Warp and kit-less execution
The `isaaclab_newton` backend is powered by Newton and MuJoCo-Warp. Isaac Lab describes kit-less operation and multiple solver paths, which can be attractive for high-throughput GPU training or teams already working with MuJoCo model semantics.
Task presets, sensors, renderers and visualization combinations still differ across backends in the current beta. Check that the selected environment has a Newton preset and that its sensors, constraints, contacts and reset behavior are supported before treating throughput results as equivalent.
Run one environment twice with a controlled record
Use the release's official environment listing and launcher rather than copying a command from an older article. Record the Git commit or release tag, backend configuration, task ID, seed, device, number of environments, timestep, solver parameters and driver/runtime information.
First run a small headless smoke test, then increase parallel environments while watching stability and memory. The comparison is valid only if the observation, action, reward, reset and evaluation definitions remain identical.
| Record | Why it matters | Pass condition |
|---|---|---|
| Release and commit | Commands and defaults change | Another user can fetch the same code |
| Task and backend config | Not every task supports both backends | Both runs use explicit supported presets |
| Physics parameters | Contact and integration affect learning | Differences are documented, not hidden |
| Seed and evaluation set | Training variance can dominate | Repeated results include uncertainty |
| Throughput and task metrics | Fast simulation can still learn poorly | Speed and policy quality are reported together |
Do not compare frames per second alone
Measure environment steps per second, GPU memory, wall-clock training time, numerical instability and held-out task success. Also inspect contact behavior and learned gait or manipulation quality, because a faster backend can converge to a policy that exploits a different approximation.
The domain randomization guide helps vary uncertain parameters, while the sim-to-real failure guide explains why randomization cannot repair an incorrect interface, missing latency or wrong actuator model.
Keep the hardware gate separate
A policy that succeeds in either backend is a simulation result. Before hardware, verify units, joint ordering, limits, control rate, communication delay, observation timing, action scaling, emergency stop and a low-energy test sequence.
Export the exact policy and normalization statistics used in evaluation, then use a staged test with conservative limits and supervision. Backend agreement can increase confidence, but it cannot certify real-world safety or performance.
Frequently asked questions
Is Isaac Lab 3 stable and generally available?
The official release is v3.0.0-beta2.patch1 as of August 7, 2026, with Isaac Sim 6.0.1 support. It remains on the beta line rather than a stable generally available release.
Is Newton 1.0 still the current Newton release?
No. Newton 1.0 was the March 2026 GA baseline, while the official repository lists v1.4.0 as the current release on the check date.
Should every Isaac Lab project switch to MuJoCo-Warp?
No. Use it when the selected task supports the Newton backend and its semantics or throughput help the project. PhysX currently retains broader Isaac Sim feature coverage for some workflows.
Official sources checked
- NVIDIA Newton 1.0 general-availability overview
- NVIDIA Isaac Lab official releases
- Newton official releases
- Newton 1.4 MuJoCo solver documentation
- NVIDIA Isaac Lab product and learning page
Last checked: August 7, 2026