Learning Algorithms — unilab.rl.algos

The RL algorithm layer lives in the unilab distribution under unilab.rl (the former unilab-rl package is merged into this repository; issue #2372, ADR-0016 RL Reintegration Into unilab.rl And Single Distribution):

  • unilab.rl.algos.appo — APPO runner, learner, staging, worker

  • unilab.rl.algos.sac / unilab.rl.algos.flash_sac — off-policy learners and runners

  • unilab.rl.runtime.offpolicy — generic off-policy runner, worker, thread budget

  • unilab.rl.algos.common — shared actor factory, networks, normalization, compile helpers

The PPO (RSL-RL) integration lives in unilab.rl.adapters.rsl_rl: RslRlVecEnvAdapter plus distributed/seed helpers, and src/unilab/scripts/train_rsl_rl.py drives upstream rsl_rl.runners.OnPolicyRunner / rsl_rl.algorithms:PPO directly.

UniLab keeps the training entrypoints (src/unilab/scripts/train_*.py), which inject environments into the RL runners through unilab.rl.contracts.env.EnvFactory; see src/unilab/base/env_factory.py for the registry-backed adapter.

All trainers conform to a single runner contract — see Runner Lifecycle.