v1.0.1

UPDATE.md

@@ -1,4 +1,7 @@
 # UPDATE
+## 20251230
+### v1.0.1
+1. quality_score到每个step时计算, 结果更加准确, 可以有效避免站立也能获得高score得分的问题 (因为平均后的score可能相对比较高, 而每一步的score都非常低)
 ## 20251229
 ### v1.0.0
 1. 完成stress pipeline的客户端, 服务端代码, 支持异步推理

robogauge/tasks/gauge/base_gauge.py

@@ -21,6 +21,7 @@ from robogauge.tasks.robots import RobotConfig
 from robogauge.tasks.gauge.base_gauge_config import BaseGaugeConfig
 from robogauge.tasks.gauge.goal_data import GoalData, VelocityGoal, PositionGoal
 from robogauge.tasks.simulator.sim_data import SimData
+from robogauge.tasks.gauge.gauge_configs.terrain_levels_config import SEARCH_LEVELS_TERRAINS
 
 from robogauge.tasks.gauge.goals import *
 from robogauge.tasks.gauge.metrics import *
@@ -149,7 +150,7 @@ class BaseGauge:
             for metric_name, quantiles in self.results[goal].items():
                 for quantile, val in quantiles.items():
                     metrics[metric_name][quantile].append(val)
-        self.results['summary'] = {}
+        self.results['summary'] = {'quality_score': {}, 'terrain_quality_score': {}}
         for metric_name, quantiles in metrics.items():
             if metric_name not in self.results['summary']:
                 self.results['summary'][metric_name] = {}
@@ -157,14 +158,19 @@ class BaseGauge:
                 mean = float(np.mean(vals))
                 std = float(np.std(vals))
                 self.results['summary'][metric_name][quantile] = f"{mean:.4f} ± {std:.4f}"
+                if metric_name == 'quality_score':
+                    tqs = mean
+                    if self.cfg.assets.terrain_name in SEARCH_LEVELS_TERRAINS:
+                        tqs = 0.09 * (self.cfg.assets.terrain_level - 1) + 0.19 * mean
+                    self.results['summary']['terrain_quality_score'][quantile] = f"{tqs:.4f} ± {std:.4f}"
 
         save_path = Path(logger.log_dir) / "results.yaml"
         self.results["terrain_name"] = self.cfg.assets.terrain_name
         self.results["terrain_level"] = self.cfg.assets.terrain_level
 
-        with open(save_path, 'w') as file:
+        with open(save_path, 'w', encoding='utf-8') as file:
-            yaml.dump(self.results, file, allow_unicode=True, sort_keys=False)
             yaml_str = yaml.dump(self.results, allow_unicode=True, sort_keys=False)
+            file.write(yaml_str)
         logger.info(
             f"""\n{'='*20} Goals and Metrics results {'='*20}\n"""
             f"""{yaml_str}"""

robogauge/tasks/gauge/base_gauge_config.py

@@ -9,6 +9,15 @@
 '''
 from robogauge.utils.config import Config
 
+QUALITY_WEIGHTS  = {  # Weights for geometric average, to calculate quality score
+    'lin_vel_err': 2,
+    'ang_vel_err': 2,
+    'dof_limits': 1,
+    'dof_power': 1,
+    'orientation_stability': 1,
+    'torque_smoothness': 1,
+}
+
 class BaseGaugeConfig(Config):
     gauge_class = 'BaseGauge'
     write_tensorboard = False  # Whether to write tensorboard logs

robogauge/tasks/gauge/goals/base_goal.py

@@ -12,6 +12,7 @@ from collections import defaultdict
 from robogauge.utils.measure import Average
 from robogauge.tasks.gauge.goal_data import GoalData
 from robogauge.tasks.simulator.sim_data import SimData
+from robogauge.tasks.gauge.base_gauge_config import QUALITY_WEIGHTS
 
 class BaseGoal:
     name = 'base_goal'
@@ -21,7 +22,8 @@ class BaseGoal:
         self.total = 0  # total tasks
         self.sub_name = None
 
-        self._goal_mean_metrics = defaultdict(list)
+        self.goal_metrics = defaultdict(list)
+        self.goal_quality_scores = []
 
     def pre_get_goal(self) -> bool:
         raise NotImplementedError
@@ -42,13 +44,19 @@ class BaseGoal:
     
     def update_metrics(self, metrics: dict):
         """ Update step metrics for the current goal."""
+        quality_score = 1.0
         for metric_name, value in metrics.items():
-            self._goal_mean_metrics[metric_name].append(value)
+            self.goal_metrics[metric_name].append(value)
+            quality_score *= min(max(1e-9, value), 1.0) ** QUALITY_WEIGHTS[metric_name]
+        quality_score = quality_score ** (1.0 / sum(QUALITY_WEIGHTS.values()))
+        self.goal_quality_scores.append(quality_score)
     
     @property
     def goal_mean_metrics(self):
         """ Get the mean metrics for the current goal. """
-        return {k: self._analysis_metrics(v) for k, v in self._goal_mean_metrics.items()}
+        result = {k: self._analysis_metrics(v) for k, v in self.goal_metrics.items()}
+        result['quality_score'] = self._analysis_metrics(self.goal_quality_scores)
+        return result
 
     @staticmethod
     def _analysis_metrics(metrics: list):

robogauge/tasks/pipeline/multi_pipeline.py

@@ -133,7 +133,7 @@ class MultiPipeline:
         """ Process results from all processes and aggregate them. """
         multi_logger.info("📊 Aggregating Results from all runs...")
 
-        summary = {'success': {}, **self.static_info, 'summary': {}, 'terrain_weighted_summary': {}, 'quality_score': {}, 'terrain_quality_score': {}}
+        summary = {'success': {}, **self.static_info, 'summary': {}, 'terrain_weighted_summary': {}}
         finish_msg = (
             f"""\n{'='*20} Run Finish Summary {'='*20}\n"""
             f"""{'Seed':^10}{'Base Mass':^15}{'Friction':^15}{'Status':^10}\n"""
@@ -152,7 +152,6 @@ class MultiPipeline:
             multi_logger.error("No results to aggregate.")
             return
         
-        quality_score, terrain_quality_score = summary['quality_score'], summary['terrain_quality_score']
         value_collections = defaultdict(lambda: defaultdict(list))
         for result in all_results:
             for goal, metrics in result['results'].items():
@@ -161,27 +160,20 @@ class MultiPipeline:
                 for metric, means in metrics.items():
                     for mean_name, mean_value in means.items():
                         value_collections[metric][mean_name].append(float(mean_value.split(' ')[0]))
-                        quality_score[mean_name] = 1
         
         for metric, means in value_collections.items():
             summary['summary'][metric] = {}
-            summary['terrain_weighted_summary'][metric] = {}
             for mean_name, values in means.items():
                 v = float(np.mean(values))
                 summary['summary'][metric][mean_name] = f"{v:.4f} ± {float(np.std(values)):.4f}"
-                twv = v
+
+            if 'quality_score' in metric: continue
+            summary['terrain_weighted_summary'][metric] = {}
+            for mean_name, values in means.items():
+                twv = float(np.mean(values))
                 if summary['terrain_name'] in SEARCH_LEVELS_TERRAINS:
                     twv = 0.09 * (summary['terrain_level'] - 1) + 0.19 * v
                 summary['terrain_weighted_summary'][metric][mean_name] = f"{twv:.4f} ± {float(np.std(values)):.4f}"
-                weight = 1
-                if metric in ['ang_vel_err', 'lin_vel_err']:
-                    weight = 2
-                quality_score[mean_name] *= min(max(1e-9, v), 1.0) ** weight
-        for mean_name in quality_score:
-            quality_score[mean_name] = quality_score[mean_name] ** (1 / 8)  # 2 + 2 + 1 * 4
-            terrain_quality_score[mean_name] = quality_score[mean_name]
-            if summary['terrain_name'] in SEARCH_LEVELS_TERRAINS:
-                terrain_quality_score[mean_name] = 0.09 * (summary['terrain_level'] - 1) + 0.19 * quality_score[mean_name]
         
         save_path = multi_logger.log_dir / "aggregated_results.yaml"
         with open(save_path, 'w') as file:

robogauge/tasks/pipeline/stress_pipeline.py

@@ -182,10 +182,12 @@ class StressPipeline:
         metric_collections = defaultdict(lambda: defaultdict(list))
         terrain_collections = defaultdict(lambda: defaultdict(list))
         zero_terrain_count = defaultdict(lambda: 0)
+        robust_score = summary['robust_score']
         for result in all_results:
             terrain_name = result['data']['terrain_name']
             terrain_level = result['level']  # None, 0, 1, ..., 10
             scores[terrain_name] = 0.0
+            robust_score[terrain_name] = {}
             key = f'{terrain_name}_{terrain_level}'
             key += f'_baseMass{result["data"]["base_mass"]}_friction{result["data"]["friction"]}'
             if terrain_level == 0:
@@ -198,7 +200,8 @@ class StressPipeline:
                 for mean_name, value_str in means.items():
                     value = float(value_str.split(' ± ')[0])
                     metric_collections[metric][mean_name].append(value)
-            for mean_name, value in result['results']['terrain_quality_score'].items():
+            for mean_name, value_str in result['results']['summary']['terrain_quality_score'].items():
+                value = float(value_str.split(' ± ')[0])
                 terrain_collections[terrain_name][mean_name].append(value)
 
         for metric, means in metric_collections.items():
@@ -207,15 +210,14 @@ class StressPipeline:
                 values.extend([0.0] * sum(zero_terrain_count.values()))  # include zero terrains
                 summary['summary'][metric][mean_name] = f"{float(np.mean(values)):.4f} ± {float(np.std(values)):.4f}"
         
-        robust_score = defaultdict(dict)
-        robust_scores = []
         for terrain_name, means in terrain_collections.items():
             for mean_name, values in means.items():
                 values.extend([0.0] * zero_terrain_count[terrain_name])  # include zero terrains
                 robust_score[terrain_name][mean_name] = float(np.mean(values))
             scores[terrain_name] = robust_score[terrain_name]['mean@50']
-            robust_scores.append(robust_score[terrain_name]['mean@50'])
+        for terrain_name in robust_score:
-        summary['robust_score'] = dict(robust_score)
+            if len(robust_score[terrain_name]) == 0:
+                robust_score[terrain_name] = None
         summary['benchmark_score'] = float(np.mean(list(scores.values())))
         scores['benchmark'] = summary['benchmark_score']
 

setup.py

@@ -2,7 +2,7 @@ from setuptools import setup, find_packages
 
 setup(
     name="robogauge",  # 包名
-    version="1.0.0",   # 版本号
+    version="1.0.1",   # 版本号
     author="Wu Tianyang", # 你的名字
     author_email="993660140@qq.com",
     description="A generic robot RL model evaluation library based on MuJoCo",