Any existing example for training a Nova Carter to navigate a custom room? (Isaac Lab RL)

[aamchitou]

Hi everyone,

I've been working on training a Nova Carter (differential drive) to navigate autonomously inside a custom room I built as a USD scene — basically reaching goal positions while avoiding walls and obstacles, using RL with Isaac Lab 2.3 and Isaac Sim 5.1.

I've already got a fair bit working (the robot drives, sees obstacles with a lidar, and learns to head toward goals), but before I keep reinventing the wheel, I wanted to ask the community:

Is there already an official tutorial or an example project that does this kind of thing — goal-reaching navigation with obstacle avoidance for a wheeled robot? I've looked at the Anymal navigation task, but it's a quadruped and has no obstacles, so it only gets me halfway.

If something like this already exists (official or community), I'd love a pointer — no need to redo it from scratch. And if it doesn't, I'd really appreciate any advice on the recommended way to approach it (lidar setup, reward design for avoidance, that kind of thing).

Thanks a lot for any help!

[ekusiadadus]

I could not find an end-to-end Nova Carter goal-navigation and obstacle-avoidance example in Isaac Lab 2.3.x.

Your version combination itself is supported: the Isaac Lab 2.3.0 compatibility table lists Isaac Sim 4.5, 5.0, and 5.1 for the 2.3.x series. I checked the complete 2.3.0 and 2.3.2 source trees, and neither contains a Carter or COMPASS task.

The closest shipped pieces are separate:

• The ANYmal-C navigation task is a hierarchical locomotion example. Its high-level policy observes base linear velocity, projected gravity, and the pose command, and it uses a pretrained low-level ANYmal policy. It is not a Carter lidar-obstacle task.
• The ray-caster tutorial shows how to configure and update a ray-caster against terrain, but it does not provide navigation rewards, obstacle observations, or a trained policy. Because it is a terrain demo, it also does not by itself establish support for sensing moving obstacles.

For a 2.3/5.1 implementation, I would therefore treat this as a custom manager-based environment: keep your existing Carter scene, add the ray/lidar ranges and relative goal to the policy observation group, expose the differential-drive action, terminate on collision and goal arrival, and add progress, arrival, collision, and clearance terms. The ANYmal task is useful for the goal-command and reward structure, while the ray-caster tutorial supplies the sensor wiring. Neither is a complete solution by itself.

For evaluation, hold out obstacle placements and start/goal layouts that were not used during training. Report goal success, collision, and timeout rates, plus path length and time to goal, and compare against a simple non-learning navigation baseline. That will show whether the policy learned transferable avoidance rather than memorizing the training room.

There is now a Carter goal-reaching configuration in COMPASS's real2sim/isaaclab_3.0 branch, and its Real2Sim workflow can be a useful architecture reference. It is not a drop-in answer here: that workflow targets an Isaac Lab 3.0.0-rc1 container with Isaac Sim 6.1 requirements and uses robot-camera RGB/depth observations rather than lidar.

This conclusion is from version-pinned source inspection; I did not run the environment in Isaac Sim. I would be interested to hear if there is a 2.3-compatible community example that is not present in the main repository.