Getting Started

Prerequisites

AGILE runs on top of Isaac Lab. All Isaac Lab requirements (Isaac Sim 5.1, NVIDIA GPU, etc.) apply. See the Isaac Lab installation guide.

Warning

AGILE requires Isaac Lab v2.3.2 specifically. Other versions may have incompatible APIs.

Installation

1. Install Isaac Lab

Follow the Isaac Lab binaries installation guide using the pre-built Isaac Sim binaries. In summary:

# 1. Download and extract Isaac Sim pre-built binaries
#    https://docs.isaacsim.omniverse.nvidia.com/latest/installation/download.html

# 2. Clone Isaac Lab and check out the required version
git clone https://github.com/isaac-sim/IsaacLab.git && cd IsaacLab
git checkout v2.3.2

# 3. Create symlink to Isaac Sim
ln -s /path/to/isaac-sim _isaac_sim

# 4. Create the conda environment
./isaaclab.sh --conda agile_env
conda activate agile_env

# 5. Install dependencies (Linux)
sudo apt install cmake build-essential

# 6. Install Isaac Lab extensions
./isaaclab.sh --install

2. Install AGILE

With the agile_env conda environment active:

conda activate agile_env
export ISAACLAB_PATH=/path/to/IsaacLab

git clone git@github.com:nvidia-isaac/WBC-AGILE.git agile && cd agile
git lfs pull

# Install AGILE + dependencies into the Isaac Lab environment
./scripts/setup/install_deps_local.sh

# Verify custom rsl_rl is installed correctly
python scripts/verify_rsl_rl.py

3. Set up pre-commit hooks (optional, for development)

./scripts/setup/setup_hooks.sh
pre-commit run --all-files  # verify

Quick Start

# Validate an environment (no trained policy needed)
python scripts/play.py --task Velocity-T1-v0 --num_envs 2

# Train a velocity tracking policy
python scripts/train.py \
    --task Velocity-T1-v0 \
    --num_envs 2048 \
    --headless \
    --logger wandb

# Evaluate a trained policy
python scripts/eval.py \
    --task Velocity-T1-v0 \
    --checkpoint /path/to/model.pt \
    --num_envs 32

See Training Guide for the full training guide, and Pre-trained Policies for available checkpoints and policy formats.

Project Structure

agile/
├── agile/
│   ├── algorithms/
│   │   ├── rsl_rl/          # Custom RSL-RL with TensorDict support
│   │   └── evaluation/      # Evaluation metrics and reporting
│   ├── data/policy/          # Pre-trained checkpoints
│   ├── isaaclab_extras/      # Isaac Lab extensions
│   ├── sim2mujoco/           # Sim-to-MuJoCo transfer
│   └── rl_env/
│       ├── assets/           # Robot USD assets
│       ├── mdp/              # MDP components (rewards, actions, observations, ...)
│       ├── tasks/            # Task definitions (self-contained configs)
│       └── rsl_rl/           # RSL-RL integration wrappers
├── scripts/                  # train.py, eval.py, play.py, export, setup
├── tests/                    # Unit and E2E tests
├── workflows/                # Docker + OSMO remote training configs
└── run.py                    # CLI for remote OSMO training/eval