rl-isaaclab

name	rl-isaaclab
description	Entry point for Moleworks IsaacLab RL work in `moleworks_ext`. Use when working on local smoke tests, Euler submits, run sync, policy playback, benchmark interpretation, or experiment tracking. Routes to narrower IsaacLab skills for cluster ops and benchmarking.

RL + IsaacLab Router

Use this as the default entry point for moleworks_ext RL work. Keep it thin and load narrower skills for the specific workflow step you need.

Repo Root

• Prefer the active moleworks_ext checkout the user is already using.
• If you need a default on this machine, start from /home/lorenzo/moleworks/moleworks_ext.
• Read repo AGENTS.md plus docs/AI_RESEARCHER_WORKFLOW.md before changing workflow or commands.

Always-Loaded Rules

• Any script that touches Isaac Sim or IsaacLab must run via /workspace/isaaclab/isaaclab.sh -p.
• Local smoke first, with W&B disabled.
• Real training runs use W&B and should be tracked in the experiment docs.
• Do not recommend checkpoints from sync alone. Require benchmark evidence.
• Prefer repo scripts and checked-in helpers over ad-hoc command reconstruction.
• Do not run long noisy training jobs yourself unless the user explicitly wants live execution and the output is tightly bounded.

Route To Narrower Skills

• For cluster submit, squeue/sacct, failed-job debugging, sync, and docs/EXPERIMENTS_* hygiene, also use rl-isaaclab-cluster-ops.
• For checkpoint benchmarking, policy playback, TensorBoard temporal plots, and run-to-run comparisons, also use rl-isaaclab-benchmark.
• If IsaacLab debugging also involves ROS parity stacks, TF, Dig3D replay, or controller-side checks, also use newton-ros-parity and ros2-debugging.
• If the task needs many code/doc searches, W&B queries, log inspection, or benchmark diffs, also use moleworks-subagent-orchestrator.

Branch References

Read only the docs that match the current task:

• docs/AI_RESEARCHER_WORKFLOW.md
• docs/EXPERIMENTS_ONGOING.md
• docs/EXPERIMENTS_RUN.md
• docs/MULTI_GPU_TRAINING.md
• docs/MULTI_GPU_BALANCED_TRAINING.md

Default Starting Points

• Train: scripts/rsl_rl/train.py
• Play: scripts/rsl_rl/play.py
• Cluster submit: docker/cluster/cluster_interface.sh
• Cluster sync: docker/cluster/sync_experiments.sh
• Live run report: scripts/utils/cluster_run_report.sh
• Config diff: scripts/utils/compare_run_configs.py
• Excavation benchmark: scripts/mole_environments/excavation3D/benchmark_excavation.py
• Temporal plots: scripts/plot_tb_scalars.py

If the task starts to sprawl across cluster ops, benchmarks, ROS parity, and repo archaeology at the same time, stop and load the narrower skills rather than expanding this one.