v1.0.5; add num_processes and support dual moe

UPDATE.md

@@ -1,7 +1,8 @@
 # UPDATE
 ## 20260106
 ### v1.0.5
-1. `robogauge/scripts/server.py`加入`--port`设置端口配置选项
+1. `robogauge/scripts/server.py`加入`--port`设置端口配置选项, 加入`--num-processes`设置并行进程数量
+2. `go2_moe`新增dual_moe模型支持, 预测返回三个对象, `student_weights, actor_weights, action`
 ## 20260103
 ### v1.0.4
 1. 加入save_additional_output参数, 启动会自动记录每个step的latent和weights

robogauge/scripts/server.py

@@ -31,7 +31,7 @@ default_args_list = [
     '--stress-benchmark',
     '--stress-terrain-names', 'flat', 'wave', 'slope_fd',  'slope_bd', 'stairs_fd', 'stairs_bd', 'obstacle',
     # '--stress-terrain-names', 'flat', 'wave',
-    '--num-processes', '30',
+    # '--num-processes', '30',  # Set in CLI
     '--seeds', '0', '1', '2',
     '--search-seeds', '0', '1', '2', '3', '4',
     '--frictions', '0.5', '0.75', '1.0', '1.25', '1.5', '1.75', '2.0', '2.25', '2.5',
@@ -95,6 +95,7 @@ def run_api_server(input_queue: multiprocessing.Queue, result_dict: dict, port=9
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument('--port', type=int, default=9973, help='API server port')
+    parser.add_argument('--num-processes', type=int, default=30, help='Number of parallel processes for StressPipeline')
     args_cli = parser.parse_args()
     print("🤖 RoboGauge Evaluation Server Starting...")
     ctx = multiprocessing.get_context('spawn')
@@ -122,7 +123,12 @@ def main():
                 results_store[task_id] = {"status": ResponseStatus.PROCESSING}
                 
                 args_list = default_args_list.copy()
-                args_list += ['--model-path', task_data.model_path, '--task-name', task_data.task_name, '--experiment-name', task_data.experiment_name]
+                args_list += [
+                    '--model-path', task_data.model_path,
+                    '--task-name', task_data.task_name,
+                    '--experiment-name', task_data.experiment_name,
+                    '--num-processes', str(args_cli.num_processes),
+                ]
                 args = parse_args(args_list)
 
                 print(f"📋 Running with args:")

robogauge/tasks/robots/go2/go2_moe.py

@@ -23,13 +23,23 @@ class Go2MoE(Go2):
     def get_action(self, obs: np.ndarray):
         obs_tensor = torch.tensor(obs, dtype=torch.float32).unsqueeze(0).to(self.device)
         action, results = self.model(obs_tensor)
-        if isinstance(results, tuple) and len(results) == 2:
-            weights, latent = results
+        if isinstance(results, tuple):
+            latent = results[-1]
             latent = latent.detach().cpu().numpy().squeeze(0) if latent is not None else None
-            weights = weights.detach().cpu().numpy().squeeze(0) if weights is not None else None
+            student_weights = None
+            actor_weights = None
+            if len(results) >= 2:
+                student_weights = results[0]
+                student_weights = student_weights.detach().cpu().numpy().squeeze(0) if student_weights is not None else None
+            if len(results) >= 3:
+                actor_weights = results[1]
+                actor_weights = actor_weights.detach().cpu().numpy().squeeze(0) if actor_weights is not None else None
             if self.cfg.control.save_additional_output:
                 self.save_info['latent'].append(latent)
-                self.save_info['weights'].append(weights)
+                if student_weights is not None:
+                    self.save_info['weights'].append(student_weights)
+                if actor_weights is not None:
+                    self.save_info['actor_weights'].append(actor_weights)
         action = action.detach().cpu().numpy().squeeze(0)[self.model2mj_idx]
         self.last_action = action
         target_dof_pos = action * self.action_scale + self.default_dof_pos