feat(training): release V0.9.1 避障训练与基础策略迁移

training_server/rl/scripts/evaluate.py

@@ -32,7 +32,6 @@ from mjlab.tasks.registry import list_tasks, load_env_cfg, load_rl_cfg, load_run
 from mjlab.tasks.velocity.mdp import UniformVelocityCommandCfg
 from mjlab.utils.torch import configure_torch_backends
 from torch.utils.tensorboard import SummaryWriter
-
 from tuning.schema import apply_reward_configuration, validate_configuration
 
 SCENARIOS = (
@@ -61,6 +60,7 @@ class EvaluateConfig:
   checkpoint: str
   output: str
   reward_config: str | None = None
+  task_config: str | None = None
   num_envs: int = 256
   steps_per_seed: int = 1000
   seeds: tuple[int, ...] = field(default_factory=lambda: (101, 202, 303))
@@ -163,31 +163,41 @@ def run_evaluation(task_id: str, cfg: EvaluateConfig) -> dict:
   checkpoint = Path(cfg.checkpoint).expanduser().resolve(strict=True)
   output = Path(cfg.output).expanduser().resolve()
   output.parent.mkdir(parents=True, exist_ok=True)
-  per_seed = [_evaluate_seed(task_id, cfg, seed) for seed in cfg.seeds]
-  metrics = {
-    key: fmean(seed_metrics[key] for seed_metrics in per_seed)
-    for key in (*METRIC_NAMES, "fall_rate")
-  }
-  deviations = {
-    key: pstdev(seed_metrics[key] for seed_metrics in per_seed)
-    for key in (*METRIC_NAMES, "fall_rate")
-  }
-  result = {
-    "protocolVersion": 1,
-    "taskId": task_id,
-    "checkpoint": checkpoint.name,
-    "checkpointSha256": _sha256(checkpoint),
-    "seeds": list(cfg.seeds),
-    "numEnvs": cfg.num_envs,
-    "stepsPerSeed": cfg.steps_per_seed,
-    "scenarios": [list(value) for value in SCENARIOS],
-    "metrics": metrics,
-    "metricStd": deviations,
-    "seedMetrics": [
-      {"seed": seed, "metrics": values}
-      for seed, values in zip(cfg.seeds, per_seed, strict=True)
-    ],
-  }
+  if task_id == "Unitree-Go2-ObstacleAvoidance":
+    from scripts.evaluate_obstacle import run_obstacle_evaluation
+    checkpoint_hash = _sha256(checkpoint)
+    result = run_obstacle_evaluation(task_id, cfg)
+    if _sha256(checkpoint) != checkpoint_hash:
+      raise ValueError("Checkpoint changed during evaluation")
+    result.update({"taskId": task_id, "checkpoint": checkpoint.name,
+                   "checkpointSha256": checkpoint_hash})
+    metrics = result["metrics"]
+  else:
+    per_seed = [_evaluate_seed(task_id, cfg, seed) for seed in cfg.seeds]
+    metrics = {
+      key: fmean(seed_metrics[key] for seed_metrics in per_seed)
+      for key in (*METRIC_NAMES, "fall_rate")
+    }
+    deviations = {
+      key: pstdev(seed_metrics[key] for seed_metrics in per_seed)
+      for key in (*METRIC_NAMES, "fall_rate")
+    }
+    result = {
+      "protocolVersion": 1,
+      "taskId": task_id,
+      "checkpoint": checkpoint.name,
+      "checkpointSha256": _sha256(checkpoint),
+      "seeds": list(cfg.seeds),
+      "numEnvs": cfg.num_envs,
+      "stepsPerSeed": cfg.steps_per_seed,
+      "scenarios": [list(value) for value in SCENARIOS],
+      "metrics": metrics,
+      "metricStd": deviations,
+      "seedMetrics": [
+        {"seed": seed, "metrics": values}
+        for seed, values in zip(cfg.seeds, per_seed, strict=True)
+      ],
+    }
   output.write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
   writer = SummaryWriter(log_dir=str(output.parent / "evaluation-events"))
   try: