prevent gradient explosion

rl_game/get_up/config/ppo_cfg.yaml

@@ -17,7 +17,7 @@ params:
           name: default
         sigma_init:
           name: const_initializer
-          val: 1.2
+          val: 0.5
         fixed_sigma: False
     mlp:
       units: [512, 256, 128]
@@ -39,21 +39,21 @@ params:
     reward_shaper:
       scale_value: 1.0
     normalize_advantage: True
-    gamma: 0.98
+    gamma: 0.96
     tau: 0.95
-    learning_rate: 5e-4
+    learning_rate: 3e-4
     lr_schedule: adaptive
-    kl_threshold: 0.008
+    kl_threshold: 0.015
     score_to_win: 20000
     max_epochs: 500
     save_best_after: 50
     save_frequency: 100
-    grad_norm: 0.5
+    grad_norm: 1.0
-    entropy_coef: 0.008
+    entropy_coef: 0.005
     truncate_grads: True
     bounds_loss_coef: 0.001
     e_clip: 0.2
-    horizon_length: 128
+    horizon_length: 65536
     minibatch_size: 8192
     mini_epochs: 4
     critic_coef: 1

rl_game/get_up/config/t1_env_cfg.py

@@ -26,27 +26,35 @@ def standing_with_feet_reward(
         force_threshold: float = 20.0,
         max_v_z: float = 0.5
 ) -> torch.Tensor:
-
+    # 增加防护：从场景中安全获取 body 索引
     head_idx, _ = env.scene["robot"].find_bodies("H2")
     pelvis_idx, _ = env.scene["robot"].find_bodies("Trunk")
 
-    curr_head_h = torch.clamp(env.scene["robot"].data.body_state_w[:, head_idx[0], 2], 0.0, 2.0)
+    # 1. 高度奖励：使用更稳定的归一化，限制范围在 [0, 1]
-    curr_pelvis_h = torch.clamp(env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2], 0.0, 2.0)
+    curr_head_h = env.scene["robot"].data.body_state_w[:, head_idx[0], 2]
+    curr_pelvis_h = env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2]
 
-    head_score = torch.tanh(curr_head_h / (min_head_height + 1e-6) * 2.0)
+    # 使用 sigmoid 或简单的 min-max 映射，避免除以极小值
-    pelvis_score = torch.tanh(curr_pelvis_h / (min_pelvis_height + 1e-6) * 2.0)
+    head_score = torch.clamp(curr_head_h / min_head_height, 0.0, 1.2)
+    pelvis_score = torch.clamp(curr_pelvis_h / min_pelvis_height, 0.0, 1.2)
     height_reward = (head_score + pelvis_score) / 2.0
 
+    # 2. 足部受力：增加对 NaN 的防御
     contact_sensor = env.scene.sensors.get(sensor_cfg.name)
+    # 某些步数传感器可能未初始化，加个判空
+    if contact_sensor is None: return torch.zeros(env.num_envs, device=env.device)
+
     foot_forces_z = torch.sum(contact_sensor.data.net_forces_w[:, :, 2], dim=-1)
+    # 对巨大的冲击力做剪裁，防止 sigmoid 输入过大
+    foot_forces_z = torch.clamp(foot_forces_z, 0.0, 500.0)
     force_weight = torch.sigmoid((foot_forces_z - force_threshold) / 5.0)
+
+    # 3. 垂直速度惩罚：使用更平滑的惩罚
     root_vel_z = env.scene["robot"].data.root_lin_vel_w[:, 2]
-    vel_penalty = torch.exp(-2.0 * torch.clamp(torch.abs(root_vel_z) - max_v_z, min=0.0))
+    vel_penalty = torch.exp(-torch.abs(root_vel_z) / max_v_z)
 
-    influence_weight = torch.clamp((curr_pelvis_h - 0.2) / 0.4, min=0.0, max=1.0)
+    # 逻辑组合：高度 * 稳定性
-    combined_reward = height_reward * ((1.0 - influence_weight) + influence_weight * force_weight * vel_penalty)
+    return height_reward * (0.5 + 0.5 * force_weight * vel_penalty)
-
-    return combined_reward
 
 
 def universal_arm_support_reward(
@@ -138,15 +146,18 @@ def is_standing_still(
 # --- 2. 配置类 ---
 
 T1_JOINT_NAMES = [
-    # 腿部
+
+    'Head_yaw', 'Head_pitch'
+                
+    'Left_Shoulder_Pitch', 'Left_Shoulder_Roll', 'Left_Elbow_Pitch', 'Left_Elbow_Yaw',
+    'Right_Shoulder_Pitch', 'Right_Shoulder_Roll', 'Right_Elbow_Pitch', 'Right_Elbow_Yaw',
+
+    'Waist'
+    
     'Left_Hip_Pitch', 'Right_Hip_Pitch', 'Left_Hip_Roll', 'Right_Hip_Roll',
     'Left_Hip_Yaw', 'Right_Hip_Yaw', 'Left_Knee_Pitch', 'Right_Knee_Pitch',
     'Left_Ankle_Pitch', 'Right_Ankle_Pitch', 'Left_Ankle_Roll', 'Right_Ankle_Roll',
-    # 手臂
+
-    'Left_Shoulder_Pitch', 'Left_Shoulder_Roll', 'Left_Elbow_Pitch', 'Left_Elbow_Yaw',
-    'Right_Shoulder_Pitch', 'Right_Shoulder_Roll', 'Right_Elbow_Pitch', 'Right_Elbow_Yaw',
-    # 腰部
-    'Waist'
 ]
 
 
@@ -180,7 +191,7 @@ class T1EventCfg:
                 "yaw": (-3.14, 3.14), # 全向旋转
                 "x": (0.0, 0.0),
                 "y": (0.0, 0.0),
-                "z": (0.1, 0.2),
+                "z": (0.3, 0.4),
             },
             "velocity_range": {},
         },
@@ -244,7 +255,7 @@ class T1GetUpRewardCfg:
     # 6. 成功终极大奖
     is_success = RewTerm(
         func=is_standing_still,
-        weight=2000.0,
+        weight=800.0,
         params={
             "min_head_height": 1.05,
             "min_pelvis_height": 0.75,