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

training_server/tests/test_server.py

@@ -10,6 +10,7 @@ from unittest.mock import patch
 
 sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 from server import (  # noqa: E402
+    DEFAULT_TASKS,
     MAX_JOBS,
     ApiError,
     TrainingJob,
@@ -79,6 +80,70 @@ out.write_bytes(b'onnx')
         with self.assertRaises(ApiError):
             self.manager.parse_config(self.payload(wandbMode="login"))
 
+    def test_custom_task_metadata_and_validated_deployment(self):
+        self.manager.tasks = DEFAULT_TASKS
+        health = self.manager.health()
+        task = next(
+            item for item in health["taskMetadata"] if item["id"] == "Unitree-Go2-ObstacleAvoidance"
+        )
+        self.assertEqual(task["sensorTypes"], ["raycast"])
+        config = self.manager.parse_config(
+            self.payload(
+                taskId=task["id"],
+                terrainPreset="discrete_obstacles",
+                terrainParams={"obstacle_count": 8, "friction": 0.9},
+                sensorCfg={"type": "raycast", "fov": 100, "maxDistance": 5},
+            )
+        )
+        self.assertEqual(config.deployment["observationSize"], 81)
+        self.assertEqual(config.deployment["terrain"]["actualObstacleCount"], 8)
+        self.assertEqual(config.deployment["sensorCfg"]["rayCount"], 32)
+        self.assertEqual(
+            TrainingJob(id="a" * 32, config=config).public()["deployment"], config.deployment
+        )
+        rough = self.manager.parse_config(self.payload(taskId="Unitree-Go2-Rough"))
+        self.assertFalse(rough.deployment["browserCompatible"])
+        self.assertEqual(rough.deployment["observationSize"], 234)
+
+    def test_custom_config_rejects_unknown_nonfinite_and_incompatible_fields(self):
+        self.manager.tasks = DEFAULT_TASKS
+        for fields in (
+            {"terrainPreset": "plane;touch /tmp/injected"},
+            {"terrainPreset": ["plane"]},
+            {"terrainParams": {"friction": float("nan")}},
+            {"terrainParams": {"size": float("inf")}},
+            {"terrainParams": {"size": 10**400}},
+            {"terrainParams": {"obstacle_count": True}},
+            {"terrainParams": {"obstacle_count": 1.5}},
+            {"terrainParams": {"mjcf": "<include/>"}},
+            {"terrainParams": {"obstacle_height_min": 1, "obstacle_height_max": 0.1}},
+            {"sensorCfg": {"rayCount": 1_000_000}},
+            {"sensorCfg": {"maxDistance": 1, "safetyDistance": 1}},
+            {"sensorCfg": {"fov": float("nan")}},
+            {"sensorType": "camera_depth"},
+            {"rewardPresetId": "a" * 32},
+        ):
+            with self.subTest(fields=fields), self.assertRaises(ApiError):
+                self.manager.parse_config(
+                    self.payload(taskId="Unitree-Go2-ObstacleAvoidance", **fields)
+                )
+        with self.assertRaises(ApiError):
+            self.manager.parse_config(self.payload(sensorCfg={"fov": 90}))
+
+    def test_custom_job_passes_server_owned_json_file_without_shell(self):
+        self.manager.tasks = DEFAULT_TASKS
+        config = self.manager.parse_config(self.payload(taskId="Unitree-Go2-ObstacleAvoidance"))
+        job = TrainingJob(id="b" * 32, config=config)
+        self.manager.jobs[job.id] = job
+        with patch("server.subprocess.Popen", wraps=subprocess.Popen) as popen:
+            self.manager._run(job)
+        args = popen.call_args.args[0]
+        self.assertNotIn("shell", popen.call_args.kwargs)
+        path = Path(args[args.index("--task-config") + 1])
+        self.assertTrue(path.is_relative_to(self.root))
+        self.assertEqual(json.loads(path.read_text()), config.task_config)
+        self.assertEqual(job.state, "succeeded")
+
     def test_builds_argument_array_without_shell(self):
         config = self.manager.parse_config(self.payload(device="gpu", gpuIds=[0, 2]))
         command = self.manager.command_for(config)
@@ -89,8 +154,13 @@ out.write_bytes(b'onnx')
 
     def test_resolves_reward_preset_to_inline_validated_trainer_argument(self):
         preset_id = "f" * 32
-        reward_config = {"weights": {"pose": 1.2}, "params": {}}
-        self.manager.preset_resolver = lambda value: reward_config if value == preset_id else None
+        from tuning.schema import base_configuration
+
+        reward_config = base_configuration()
+        reward_config["weights"]["pose"] = 1.2
+        self.manager.preset_resolver = lambda value, task: (
+            reward_config if value == preset_id and task == "Unitree-Go2-Flat" else None
+        )
         config = self.manager.parse_config(self.payload(rewardPresetId=preset_id))
         command = self.manager.command_for(config)
         index = command.index("--reward-config-json")