Install Support for Reinforcement Learning
RLBotics is a lightweight, GPU-accelerated PyTorch library for reinforcement learning. It supports multi-environment training across Gymnasium, mjlab and Isaac Lab, ONNX export, and deployment with NumPy alone.
It ships as its own package, telekinesis-rlbotics, open source under Apache 2.0 on GitHub. It is independent of the telekinesis-ai SDK, so it needs neither an API key nor the SDK — just its own environment and a PyTorch build.
Check the requirements
| Requirement | Notes |
|---|---|
| Python 3.10–3.12 | Isaac Lab needs exactly 3.11 |
| PyTorch | Installed automatically. Install it first only to pin a specific CUDA toolkit |
| NVIDIA GPU | Required for mjlab and Isaac Lab training, not for Gymnasium |
To match a specific CUDA toolkit, install PyTorch before RLBotics:
pip install torch --index-url https://download.pytorch.org/whl/cu128Install RLBotics
pip install telekinesis-rlboticsThat gives you the library itself: PPO, the runner, the models, logging, checkpointing and ONNX export — plus onnxruntime to run an exported policy, imageio for video and pyyaml for the configs, which are all base dependencies rather than extras. It is all you need to train against your own custom environment.
Add a simulator backend
A simulator comes from an extra, and more than one can be installed.
# Gymnasium classic-control and MuJoCo tasks.
# Runs on macOS, Linux and Windows. No GPU required.
pip install "telekinesis-rlbotics[gym]"# mjlab tasks on MuJoCo Warp, thousands of parallel environments.
# Needs Linux or Windows with an NVIDIA GPU.
pip install "telekinesis-rlbotics[mjlab]"
# mjlab picks its physics backend through its own extras, so this is a second step.
# Match the CUDA version to your toolkit. On macOS use "mjlab[cpu]" (evaluation only).
pip install "mjlab[cu128]"# Isaac Lab tasks on Isaac Sim, the largest task library of the three.
# Needs Python 3.11, Linux (GLIBC 2.35+) or Windows, and an NVIDIA GPU.
# No macOS build of Isaac Sim exists.
#
# The Isaac Sim wheels are hosted by NVIDIA rather than PyPI, so this extra
# only resolves with their index.
pip install "telekinesis-rlbotics[isaaclab]" --extra-index-url https://pypi.nvidia.com# ruff, pylint and pytest, for working on RLBotics itself. Runs anywhere.
pip install "telekinesis-rlbotics[dev]"Train from a configuration
A run is described by one YAML file, and one script trains any of them — the config's env.framework decides which simulator is used. Both the scripts and the configs live in the repository rather than in the wheel, so clone it:
git clone https://github.com/telekinesis-ai/telekinesis-rlbotics.git
cd telekinesis-rlbotics# A pendulum swing-up, solved in about two minutes on CPU
python examples/training_example.py configs/gymnasium/Pendulum-v1.yaml
# Or a MuJoCo humanoid learning to walk, roughly 20 minutes
python examples/training_example.py configs/gymnasium/Humanoid-v5.yamlpython examples/training_example.py configs/mjlab/Mjlab-Velocity-Flat-Unitree-G1.yamlpython examples/training_example.py configs/isaaclab/Isaac-Velocity-Flat-Anymal-C-v0.yamlEach run ends by exporting policy.onnx and acting with it, which exercises the whole path — training, checkpointing, export and deployment. Watch the curves with:
tensorboard --logdir logsMore tasks are in configs/<framework>/, and configs/example.yaml is an annotated reference covering every field. See Configuration for the schema.
Next Steps
Install BabyROS
Install the lightweight pub/sub and client/server middleware for distributed robotics communication.
Next →