Policy Training
Policy Training
rlbotics trains PPO policies in simulation - pick a simulator backend, run it, and watch the curves in TensorBoard.
TIP
Install with pip install telekinesis-rlbotics — see Install Support for Reinforcement Learning for the simulator backends.
Training Backends
Three simulators, all driven through the same runner. Gymnasium runs anywhere with no GPU and is the one to start with; mjlab and Isaac Lab trade a handful of CPU environments for thousands of parallel ones on the GPU.
Train in Gymnasium
Trains on any continuous control task from its id alone. Runs on macOS, Linux and Windows without a GPU.
View →Train in mjlab
Trains on MuJoCo Warp across thousands of parallel GPU environments, with 12 registered locomotion and manipulation tasks.
View →Train in Isaac Lab
Trains on Isaac Sim, the largest task library of the three, covering quadrupeds, humanoids and Franka manipulation.
View →Configuring a Run
One typed config tree describes a training run, and every config validates itself on construction.
Where to Go Next?
Continue to the next tutorial.
Policy Tuning and Deployment
Read the training curves, tune what matters, and export a policy that runs with numpy alone.
Next tutorial →