Language Language English 简体中文

unilab.algos.torch.appo.runner

APPO Runner — Asynchronous PPO with native multiprocessing.

Pipeline:

1. Collector subprocess publishes rollout payloads → RolloutRingBuffer

2. Learner reads rollouts, computes V-trace corrected updates

3. Weights synced back to collector via SharedWeightSync

Classes

APPORunner

APPO async runner using shared memory.

class unilab.algos.torch.appo.runner.APPORunner

Bases: AsyncRunner

APPO async runner using shared memory.

Parameters:

• env_name (str)

• env_cfg_overrides (dict)

• rl_cfg (dict)

• device (str | None)

• collector_device (str | None)

• sim_backend (str)

• num_envs (int)

• steps_per_env (int)

• num_workers (int)

• replay_queue_size (int)

• seed (int | None)

• resume_path (str | None)

__init__(env_name, env_cfg_overrides, rl_cfg, device=None, collector_device=None, sim_backend='mujoco', num_envs=1024, steps_per_env=24, num_workers=1, replay_queue_size=3, seed=None, resume_path=None)

Parameters:

• env_name (str)

• env_cfg_overrides (dict)

• rl_cfg (dict)

• device (str | None)

• collector_device (str | None)

• sim_backend (str)

• num_envs (int)

• steps_per_env (int)

• num_workers (int)

• replay_queue_size (int)

• seed (int | None)

• resume_path (str | None)

learn(max_iterations=1500, save_interval=50, log_dir='logs', logger_type='tensorboard')

Parameters:

• max_iterations (int)

• save_interval (int)

• log_dir (str)

• logger_type (str)

Return type:

None