Isaac Lab

isaac-lab

The GPU-parallel RL/IL training layer of robium's NVIDIA stack, built on top of an already-running Isaac Sim: prebuilt environments and tasks (isaaclab_tasks), training entry points for several RL libraries, an imitation-learning path for generating and training on simulated demonstrations, and exporting a trained policy. Isaac Lab (isaac-sim/IsaacLab, current release v3.0.0-beta2.patch1, published 2026-07-02 — verified via direct fetch of the GitHub releases API this session) is NVIDIA's own framework layered on Isaac Sim, not a separate product to install independently. Its main branch's own installation docs state support for Isaac Sim 4.5/5.0/5.1 and recommend the latest 5.1.0 release specifically (verified via direct fetch of the installation docs on 2026-07-10) — that may trail the newest Isaac Sim release the isaac-sim skill cites, so confirm the current supported-version pairing before installing rather than assuming the two always track together.

When to use this skill

• Running a prebuilt Isaac Lab task, training a policy with an RL library on top of a working Isaac Sim install, generating/training on simulated demonstrations, or exporting a trained policy for deployment.
• The trigger phrases in the description: 'isaac lab', 'GPU RL for robots', 'train in isaac', sim-to-real policy training in the NVIDIA stack.
• Cross-references — go to the sibling skill instead when the question is:
  • Isaac Sim itself is not installed/working yet (GPU floor, container, USD scene, robot/sensor import, ROS 2 bridge) → isaac-sim. This skill assumes Isaac Sim is already running; it only adds the training layer on top.
  • Imitation learning on datasets recorded from a real robot (the LeRobotDataset format, lerobot-train/lerobot-record) → lerobot. This skill's own imitation-learning path (see Usage patterns) starts from demonstrations recorded inside Isaac Sim, not real hardware — that distinction is the actual boundary, not "imitation learning" as a category.
  • Whether to use the NVIDIA stack (Isaac Sim/Lab) at all vs. LeRobot's own sim/eval tooling → architect decides this, gated on the GPU floor (see Platform gotchas).
  • Which simulator to use in general, before Isaac Sim is chosen → simulation.
  • Deciding data-sourcing strategy (how much sim-generated vs. real data a project needs) → the data umbrella skill. This skill only covers the mechanics of Isaac Lab's own demonstration-generation and training tools, not the sourcing decision.

Key directives

• Delegation posture: embed + links. The install-on-top-of-Isaac-Sim sequence, task-ID convention, and the RL/IL/export commands below are embedded because no single upstream page walks a new robium project through all three together — but every command is sourced from isaac-sim.github.io/IsaacLab's own docs or the isaac-sim/IsaacLab GitHub repo, fetched directly on 2026-07-10, rather than retyped from an older Isaac Lab release's memory. See References.
• On a cloud GPU, prefer the prebuilt Isaac Lab image over a source install. nvcr.io/nvidia/isaac-lab (latest tag observed 3.0.0-beta2-post1 on the NGC catalog 2026-07-26..28) bundles a matched Isaac Sim + Isaac Lab in one container, which sidesteps the version- pairing trap (below) and the multi-step source install. This is the battle-tested path from the go2-locomotion app on RunPod; the pip+source route in Quick start remains valid for a workstation you own. See references/prebuilt-image-runpod.md for the provisioning specifics (NGC auth, EULA env vars, the /workspace volume-shadow gotcha, and the entrypoint override), and the environments skill's GPU-and-remote reference for the general pod networking/SSH mechanics this builds on.
• The GPU/driver floor is isaac-sim's, not restated here. Isaac Lab runs inside Isaac Sim, so it inherits that skill's GPU requirement verbatim — check the exact minimum/recommended GPU and VRAM numbers there, don't re-derive or re-type them in this skill. Isaac Lab's own RL training workloads (many parallel environments) also want more VRAM headroom than a bare Isaac Sim scene; treat isaac-sim's stated floor as a minimum, not a comfortable working point for large --num_envs runs.
• Start from a prebuilt task before writing a custom environment. List and run an existing task first (see Quick start) to confirm the install works end to end with zero environment-authoring risk, the same "validate the pipeline before customizing" posture lerobot takes with a pretrained policy.
• Never write task IDs, script paths, or CLI flags from memory. Isaac Lab's task registry and script layout change across releases (the top-level scripts directory was itself reorganized into reinforcement_learning and imitation_learning subdirectories) — list the currently-registered tasks instead of assuming a task name from a prior release still exists, and re-verify script paths against isaac-sim/IsaacLab's main branch before repeating one in a real project.

Quick start

Source: isaac-sim.github.io/IsaacLab's installation and quickstart docs, and the isaac-sim/IsaacLab GitHub repo's scripts directory tree, fetched directly on 2026-07-10.

1. Confirm the GPU floor — see the isaac-sim skill; do not proceed until the target machine meets it.

2a. Cloud GPU (preferred): pull the prebuilt Isaac Lab image. nvcr.io/nvidia/isaac-lab bundles a matched Isaac Sim + Isaac Lab, so there is no separate Isaac Sim install and no version-pairing to reconcile. See references/prebuilt-image-runpod.md for the RunPod provisioning specifics (NGC auth, EULA env vars, the /workspace volume-shadow gotcha, the entrypoint override, and the observed image tag/driver/Python versions).

2b. Workstation you own (alternative): install Isaac Sim via pip, then Isaac Lab from source on top of it (per-release version pins matter — verify the current recommended Isaac Sim version against the installation docs before pinning it):

pip install "isaacsim[all,extscache]==5.1.0" --extra-index-url https://pypi.nvidia.com
git clone https://github.com/isaac-sim/IsaacLab.git --branch main
cd IsaacLab
./isaaclab.sh --install

3. List the registered tasks:

python scripts/environments/list_envs.py

4. Train on a prebuilt task with one of the shipped RL libraries (rsl_rl, skrl, rl_games, sb3):

python scripts/reinforcement_learning/skrl/train.py --task=Isaac-Ant-v0 --headless

5. Watch progress and evaluate/export — see Usage patterns.

Usage patterns

Run a prebuilt task. Task IDs follow Isaac-<Name>-v0 (manager-based workflow) or Isaac-<Name>-Direct-v0 (direct workflow) — list_envs.py (Quick start) prints the full current table with entry points, rather than guessing a name from a tutorial. --num_envs=<n> sets how many parallel environments run (the GPU-parallel core of Isaac Lab's speed advantage); drop --headless only for local interactive debugging on a machine with a display, since it costs render throughput.

Train + monitor. Each RL library ships its own train.py under its own subdirectory of the reinforcement-learning scripts tree, with a matching play.py for evaluation and checkpoint loading:

python scripts/reinforcement_learning/rsl_rl/train.py --task=Isaac-Cartpole-v0 --headless --num_envs=4096

Runs log to a timestamped directory under logs/<library>/<task>/; RSL-RL's own agent config exposes a logger field (tensorboard by default, or wandb/neptune — confirmed via direct fetch of isaaclab_rl's RL-library config on 2026-07-10) — point tensorboard --logdir logs/rsl_rl at the run directory to watch reward/loss curves live. --max_iterations overrides the task's default training length for a short smoke run before committing to a full one, the same small-scale-first posture lerobot uses for fine-tunes.

Verified Go2 locomotion walkthrough. A battle-tested RSL-RL run on the Unitree Go2 (task Isaac-Velocity-Flat-Unitree-Go2-v0, PPO) — smoke, full-training profile, reward/cost config, custom-robot scaffolding, and checkpoint portability — lives in references/go2-rl-workflow.md. One trap worth stating up front: RSL-RL writes checkpoints under the experiment name, logs/rsl_rl/unitree_go2_flat/<timestamp>/, not the task ID, so a smoke test asserting on a task-ID-shaped path fails even when training succeeded. Reward and cost also live in config, not code (a cost is a reward term with a negative weight) — the reference names the exact files and the 3-layer weight-override chain.

Evaluate and export a trained policy. play.py (same per-library directory as train.py) loads a checkpoint and runs it in the environment; for RSL-RL specifically, it also exports the policy to both TorchScript (JIT) and ONNX under the checkpoint's exported/ directory automatically — confirmed by direct fetch of the RSL-RL play.py source on 2026-07-10, which calls export_policy_to_jit/export_policy_to_onnx (or the older export_policy_as_jit/export_policy_as_onnx helpers on RSL-RL < 4.0). This exported artifact is the sim-to-real hand-off point — deploying it onto real hardware is outside this skill's depth once exported.

Imitation learning from simulated demonstrations. A separate imitation_learning/ script tree (isaaclab_mimic, robomimic, and a record_demos.py/replay_demos.py pair under the tools scripts directory) records teleoperated or scripted demonstrations inside Isaac Sim and trains a policy on them — this is the sim-side imitation-learning path, distinct from lerobot's real-robot-dataset training (see When to use this skill). Treat this as a pointer, not a full walkthrough — verify the current CLI against the imitation_learning/ and tools/ directories before running it.

Hand-off from LeRobot. lerobot-eval --env.type=isaaclab_arena loads Isaac Lab Arena through LeRobot's EnvHub mechanism (lerobot.envs.make_env) rather than this skill's own scripts — that's lerobot's territory calling into an Isaac Lab environment, not the reverse; see the lerobot skill's eval-and-sim reference for that specific invocation.

Platform gotchas

• GPU floor is isaac-sim's — don't re-derive it. No macOS, RTX-class NVIDIA GPU required; see that skill for the exact minimum/recommended numbers and how they were verified.
• Isaac Sim/Isaac Lab version pairing is narrower than "whatever's newest." Isaac Lab's main branch supports a specific Isaac Sim version window (4.5/5.0/5.1 as of 2026-07-10, recommending 5.1.0) rather than every Isaac Sim release — installing the two independently without checking this pairing is a common source of import-time failures. Re-check the installation docs' compatibility statement before pinning versions in a real project.
• Headless is the default for real training runs, same as isaac-sim. --headless avoids paying render cost for a GUI viewport during a training run with thousands of parallel environments; reserve the non-headless mode for short interactive checks on a machine with a display, per isaac-sim's own headless-first guidance.

Customization

• Different task or robot: list_envs.py (Quick start) is the source of truth for what's currently registered — pick an existing task close to the target robot/behavior before authoring a new one. When you do author one, a custom robot/task is an external project, not a fork of Isaac Lab: ./isaaclab.sh --new scaffolds a standalone repo that pip-installs Isaac Lab and gym.registers the task. The in-repo "internal task" path is only for upstreaming and is auto-disabled whenever Isaac Lab is pip-installed (i.e. inside the prebuilt NGC container), so external-project is the only path there. See references/go2-rl-workflow.md; Isaac Lab's own tutorials (linked in References) cover task authoring in depth this skill does not duplicate.
• Different RL library: swap which library's subdirectory of the reinforcement-learning scripts tree you invoke (rsl_rl, skrl, rl_games, sb3) — each wraps the same underlying Isaac Lab environment with that library's own agent config and CLI flags, so a task that works under one library isn't a guaranteed drop-in for another's config shape.
• No local GPU meeting the floor: don't try to run Isaac Lab without it — route to lerobot's own sim/eval tooling (per architect's manipulation-vertical guidance) or provision a remote GPU host meeting isaac-sim's floor first.

References

• references/prebuilt-image-runpod.md — the prebuilt nvcr.io/nvidia/ isaac-lab image (observed tag 3.0.0-beta2-post1) and its RunPod provisioning specifics: NGC auth, EULA env vars, the /workspace volume-shadow gotcha, and the entrypoint override + stop→start. Points to the environments skill for the general pod networking/SSH mechanics.
• references/go2-rl-workflow.md — the verified Unitree Go2 RSL-RL run: task IDs, smoke and full-training profile, the experiment-name log-path trap, reward/cost-in-config layout and 3-layer override chain, the external-project route for custom robots, and checkpoint/script portability on a fresh pod.
• Upstream: Isaac Lab documentation (installation, quickstart, and task/training concepts — primary source for this skill, fetched directly on 2026-07-10), isaac-sim/IsaacLab GitHub repo (the reinforcement-learning, imitation-learning, tools, and environments scripts subdirectories, fetched directly via the GitHub Contents API and raw file URLs this session — source of the exact script paths, task-ID convention, and export-format claims above). Sibling skills: isaac-sim (GPU floor, install, and the Isaac Sim instance this skill runs on top of), lerobot (alternative manipulation ML path; owns real-robot-dataset imitation learning and the isaaclab_arena EnvHub hand-off), data (data-sourcing strategy, including how much this skill's own demo-generation tools should contribute), simulation (simulator selection before Isaac Sim is chosen), architect (routes here, GPU-gated, decides isaac-lab vs. lerobot).

Changelog

• 1.1.1 (2026-08-01): anchor IDs added to claim-bearing items (learning-engine Phase 1); no content changes.
• 1.1.0 (2026-07-31): hardened from the go2-locomotion RunPod L4 run (2026-07-26..28) — added the prebuilt-image-runpod and go2-rl-workflow references, made the prebuilt nvcr.io/nvidia/isaac-lab image the preferred cloud path, and surfaced the experiment-name log-path trap and external-project custom-robot route.
• 1.0.1 (2026-07-12): skill-refiner run 1 — provenance claims date-stamped ('this session' → 2026-07-10, the authoring session) so the staleness sweep can age them.