feat(training): upgrade Isaac Lab runtime to 3.0

[agreaves-ms]

Component

Training Workflows

Problem Statement

RL training and SIL evaluation run on the digest-pinned nvcr.io/nvidia/isaac-lab:2.3.2 image. That image runs Python 3.11, so training/rl pins Python 3.11 and, for Isaac Sim ABI compatibility, NumPy 1.26.4. As a result:

• training/rl stays below the repository's Python 3.12 minimum from build: standardize on Python 3.12 as repository minimum #540.
• Dependabot ignores NumPy 2.x updates for training/rl (build: dependabot: guard /training/rl numpy + packaging against Isaac Sim ABI-breaking bumps #1029).
• There is no supported path to the Isaac Lab 3.0 runtime, which uses Python 3.12. The image default, the workflow pins, and the RL lock have to change together.

Proposed Solution

Move the shared Isaac Lab runtime to the digest-pinned nvcr.io/nvidia/isaac-lab:3.0.0-beta2-post1 image:

• Point DEFAULT_ISAAC_LAB_IMAGE, the Azure ML environment pins, and the direct OSMO workflow fallbacks at the new image.
• Move training/rl to Python 3.12 and NumPy 2.5.1, regenerate training/rl/uv.lock, and update the Dependabot NumPy guard and the setup-dev Isaac Lab commit pin.
• Keep the image-provided Torch and CUDA packages when installing locked RL dependencies. Install into a writable overlay when the interpreter's site-packages are read-only, and report runtime provenance from that overlay.

Alternatives Considered

• Stay on Isaac Lab 2.3.2 until a stable 3.0 release ships. This keeps training/rl on Python 3.11 and NumPy 1.x and leaves its NumPy updates blocked.
• Offer 3.0 only through a DEFAULT_ISAAC_LAB_IMAGE override. The RL lock resolves for a single Python version and NumPy ABI, so the image default and training/rl/uv.lock must move together.

Additional Context

Consideration	Impact
Breaking change	Isaac Lab-based RL training and SIL evaluation jobs default to Isaac Lab 3.0 with Python 3.12, and training/rl moves to NumPy 2.5.1. Custom tasks or extensions built for Isaac Lab 2.3.2 may need updates.
Validation	The target is a beta tag and hosted CI cannot run Isaac Sim, so the change needs an RL training smoke run on a GPU node.
Azure services	None added
Network modes	Unaffected
Personas	ML engineers and robotics developers running RL training or SIL evaluation

Related: #114 tracks a container upgrade that removes the simulation_shutdown.py workaround. This issue keeps that workaround in place.