RLBotics: Reinforcement Learning Skills
SUMMARY
RLBotics is a module in the Telekinesis Agentic OS for reinforcement learning (RL) skills in robotics: a lightweight, GPU-accelerated PyTorch library that trains across Gymnasium, mjlab and Isaac Lab through one pipeline, exports to ONNX, and deploys with NumPy alone. A run is described by a single YAML file, so training a different task means pointing at a different config.
Open source under Apache 2.0 — telekinesis-rlbotics on PyPI, source on GitHub.
When to Use RLBotics?
Use RLBotics for robotics applications that require learned control and behaviors, such as:
- Training RL policies for locomotion, manipulation, or control in simulation
- Simulating and validating policies before deployment
- Deploying policies sim-to-sim (e.g., from one simulator to another) or sim-to-real (from simulation to real robots)
- Integrating learned control into Physical AI pipelines alongside perception and planning
What Does RLBotics Provide?
RLBotics includes a collection of modular skills for:
- Training RL policies in simulation with support for common algorithms and backends
- Running and debugging policies in simulation (stress-testing, tuning, validation)
- Sim-to-sim deployment: running the same policy across different simulators
- Sim-to-real deployment: transferring policies from simulation to real robots with a unified interface
Guides
Train a PPO policy on a simulator, watch the curves in TensorBoard, export it to a single ONNX file, and run it on the robot with numpy alone.
Train in Gymnasium
Runs anywhere, no GPU needed. Any continuous control task works from its id alone - two minutes for Pendulum, twenty for a MuJoCo humanoid.
Start here →Train in mjlab
MuJoCo Warp on the GPU, thousands of parallel environments, and 12 registered locomotion and manipulation tasks.
Explore →Train in Isaac Lab
Isaac Sim, the largest task library of the three backends, with quadruped, humanoid and Franka manipulation tasks.
Explore →Deploy a Policy
One self-contained policy.onnx carrying the observation normalization and action scaling. Runs with numpy and onnxruntime alone.
Explore →Configuration
Every option with its default: observation groups, PPO hyperparameters, models, logging, checkpoints and resume.
Explore →Tuning Best Practices
Which curve to read first, the measured per-task recipes, why exploration collapses without a std floor, and symmetry.
Explore →Architecture
How the pieces fit: the runner owns the loop, the algorithm owns the update, an adapter owns the simulator, one file owns deployment.
Read first →Introduce a Custom Environment
Implement the VecEnv interface and the runner trains against your own simulator or robot.
Explore →Introduce a Custom Algorithm
The protocol the runner drives, how to subclass PPO, and how to add your own hyperparameters.
Explore →