LIRRA  1.0.0
Loading...
Searching...
No Matches
rl_stl.policies.ppo_checkpoint Namespace Reference

Classes

class  NetworkConfig
 
class  PPOConfig
 
class  TrainConfig
 

Functions

 load_ppo_run (str|Path run_dir)
 

Detailed Description

Config schema of ``experiments/skydio_x2/train_ppo.py`` runs and a loader for their policies.

A run directory holds ``config.yaml`` (the resolved :class:`TrainConfig`) and
``policy_params.pkl`` (Brax ``(normalizer, policy, value)`` params). Reloading the
config requires the exact dataclass shape it was written with, so the schema lives
here and the training script imports it.

Importing this module imports JAX (through the environment config), so call
:func:`rl_stl.utils.device.configure_jax_device` first.

Function Documentation

◆ load_ppo_run()

rl_stl.policies.ppo_checkpoint.load_ppo_run ( str | Path  run_dir)
Return ``(cfg, env, policy)`` for a trained run.

``env`` is the :class:`~rl_stl.envs.skydio_x2.hover.X2Hover` built from the run's own
env config (so the observation layout matches the network); ``policy(obs, key)`` is the
deterministic Brax inference function (not jitted).