| """Score the selected policy without changing its observations or actions.""" |
| from dataclasses import replace |
| from pathlib import Path |
|
|
| from .rl.config import load_rl_config |
| from .rl.service import RLService |
|
|
|
|
| def start_policy_evaluation(runner, *, mocap_zmq_endpoint=None): |
| root = Path(__file__).resolve().parents[1] |
| config = load_rl_config(runner._eval_config or root / "configs/policy_eval.yaml") |
| if mocap_zmq_endpoint is not None: |
| config = replace(config, zmq_endpoint=mocap_zmq_endpoint) |
| policy_path = Path(runner.settings.policy_path).resolve() |
| print(f"[POLICY EVAL] policy={policy_path} obs={runner.policy.obs_dim} " |
| "recording only; no training/exploration/action changes; manual stop", flush=True) |
| return RLService(config, runner.contract, runner.spec, runner.robot_base_origin_world, |
| mode="baseline", observe_only=True, observation_dim=runner.policy.obs_dim, |
| project_root=root, run_dir=runner._eval_run_dir, |
| policy_metadata={"path": str(policy_path), "name": policy_path.name, |
| "version": runner.settings.deployment_version, |
| "reference": type(runner.reference).__name__, |
| "scoring_config": str(Path(runner._eval_config or root / "configs/policy_eval.yaml").resolve())}) |
|
|
