Skip to main content

Installation and Setup

Install SuperDex Gym​

SuperDex Gym (superdex-lab) is installed as part of Project SuperDex. Its package dependencies include SuperDex Physics (superdex-physics) for simulation and SuperDex Robotics (superdex-robotics), which every environment imports at module load. Ray/RLlib is an optional training-stack dependency for apps/rllib and is installed separately.

Before installing, follow the Project SuperDex build instructions to set up the required toolchain. Use the project_superdex root as your working directory.

Run the core install:

uv sync --extra core

This command builds physics, robotics and Gym together. The README's extras table covers the other combinations.

FP64 native simulation

Build the fp64 extra and set SUPERDEX_PRECISION=fp64 before importing SuperDex to run the native Physics and Robotics simulation in FP64. SuperDex Lab has no separate FP64 package: its shipped Gym environments retain float32 observation and action spaces. This is FP64 native simulation under the existing float32 Gym interface, not an end-to-end float64 Lab API.

Verifying the Install​

Complete the core install first. Use the project_superdex root as your working directory, then run:

uv run python -c "import superdex.physics; print(superdex.physics.__file__)"
uv run python -c "import superdex.robotics; print('ok')"
uv run python -c "import superdex.lab.gym; print('ok')"
uv run python -c "from superdex.physics.paths import get_assets_root; print(get_assets_root())"

The last command prints the assets/ directory of your checkout.

Run an Environment​

Complete the install verification first. From the project_superdex root, run this headless smoke test:

On macOS, Linux, or PowerShell:

uv run python -c "
from superdex.lab.gym.envs.benchmarks.cartpole_env import CartPoleEnv, CartPoleEnvCfg

env = CartPoleEnv(CartPoleEnvCfg(render_mode=None))
try:
    env.reset()
    env.step(env.action_space.sample())
finally:
    env.close()
print('Smoke test passed')
"

With Windows Command Prompt:

uv run python -c "from superdex.lab.gym.envs.benchmarks.cartpole_env import CartPoleEnv, CartPoleEnvCfg; env = CartPoleEnv(CartPoleEnvCfg(render_mode=None)); exec('try:\n    env.reset()\n    env.step(env.action_space.sample())\nfinally:\n    env.close()'); print('Smoke test passed')"

See Run Example Environments for the interactive and video-capable scripts.

Understand Asset Discovery​

Environments resolve their scenes and meshes through superdex.physics.paths.get_assets_root() in this order:

  1. Precedence: It uses SUPERDEX_ASSETS_PATH first when that variable points at a readable directory.
  2. Fallback: Otherwise, it walks upward from the installed superdex.physics module, not from the current working directory, looking for an assets/ directory in the same source workspace.
  3. Editable install: An editable source-workspace install resolves the assets automatically through that fallback.
  4. Wheel install: A wheel installed outside the checkout does not resolve a separate clone automatically.
  5. Corrective action: When using a wheel with a separate clone, set SUPERDEX_ASSETS_PATH=<path-to-project_superdex>/assets.

Dependencies for apps/​

Complete uv sync --extra core first. It installs superdex-lab and the rest of the workspace lock, including polyscope, imageio, imageio-ffmpeg, pillow and psutil.

The scripts under apps/ have additional requirements. The apps/envs benchmarks require tqdm>=4.67.1. The apps/rllib training and video scripts require torch==2.7.1, ray[rllib]==2.49.0, moviepy, pillow>=10.1 and tensorboard. Use the project_superdex root as your working directory, then run:

# apps/envs benchmarks
uv pip install "tqdm>=4.67.1"

# apps/rllib training and video
uv pip install torch==2.7.1 --extra-index-url https://download.pytorch.org/whl/cpu
uv pip install "ray[rllib]==2.49.0" "moviepy" "pillow>=10.1" "tensorboard"

These manual installs are not recorded as requested dependencies of the root project, so running uv sync again can remove unlisted packages or restore locked versions. The superdex_lab project is unmanaged by uv, so plain uv run from its directory tree uses the existing environment without synchronizing it.

The scripts under apps/rllib/ import their siblings by bare name. Start in the project_superdex root; the first command changes the working directory to superdex_lab/apps/rllib.

cd superdex_lab/apps/rllib
uv run python train_samples.py --help

Running the Tests​

Complete the install first. The tests require pytest. Use the superdex_lab project root, the directory holding pyproject.toml and test/, as your working directory.

Install the test requirement:

uv pip install pytest

Then run the test suite from the same directory:

uv run python -m pytest test

Expect 51 passed, 3 skipped. The three skips are the test_batched_stepping_* cases, which need an agent_pose observation that none of the shipped environments has.

Next Steps​