Telekinesis - RLbotics

Telekinesis RLbotics is a lightweight, GPU-accelerated PyTorch library for Reinforcement Learning. It supports multi-environment training across Gymnasium, mjlab, and Isaac Lab, common learning algorithms, ONNX export, and deployment with NumPy alone.

Open source under Apache 2.0.

Full documentation: Telekinesis Agentic OS: RLbotics.

Requirements

• Python 3.10–3.12
• PyTorch. To match a specific CUDA toolkit, install it first:

pip install torch --index-url https://download.pytorch.org/whl/cu128

Quickstart

Gymnasium - (requirements: any OS, no GPU)

pip install "telekinesis-rlbotics[gym]"
python examples/training_example.py configs/gymnasium/Humanoid-v5.yaml

mjlab - (requirements: Linux/Windows, NVIDIA GPU)

pip install "telekinesis-rlbotics[mjlab]" "mjlab[cu128]"
python examples/training_example.py configs/mjlab/Mjlab-Velocity-Flat-Unitree-G1.yaml

Isaac Lab - (requirements: Linux/Windows, NVIDIA GPU, Python 3.11)

pip install "telekinesis-rlbotics[isaaclab]" --extra-index-url https://pypi.nvidia.com
python examples/training_example.py configs/isaaclab/Isaac-Velocity-Flat-Anymal-C-v0.yaml

tensorboard --logdir logs to watch any of them learn. More tasks in configs/<framework>/.

Options

python examples/training_example.py --help for details.

Deployment

Training exports the best checkpoint to one self-contained policy.onnx. Run it with just numpy and onnxruntime:

from telekinesis.rlbotics.policy import Policy

policy = Policy("logs/gymnasium_ppo/2026-08-06_18-08-47/policy.onnx")
action = policy.get_action(observation)   # (obs_dim,) -> (num_actions,)

Documentation

Find the documentation for RLBotics at: Telekinesis Agentic OS: RLbotics

Citation

@software{telekinesis_rlbotics,
  author = {Telekinesis GmbH},
  title  = {Telekinesis-Rlbotics: Reinforcement Learning for Robotics},
  year   = {2026},
  url    = {https://github.com/telekinesis-ai/telekinesis-rlbotics},
  note   = {Apache-2.0}
}