change model

rl_game/get_up/config/ppo_cfg.yaml

@@ -41,11 +41,11 @@ params:
     normalize_advantage: True
     gamma: 0.96
     tau: 0.95
-    learning_rate: 5e-4
+    learning_rate: 3e-4
     lr_schedule: adaptive
     kl_threshold: 0.008
     score_to_win: 20000
-    max_epochs: 1000000
+    max_epochs: 200
     save_best_after: 50
     save_frequency: 100
     grad_norm: 0.5

rl_game/get_up/config/t1_env_cfg.py

@@ -52,10 +52,9 @@ def standing_with_feet_reward(
 def arm_push_up_reward(
         env: ManagerBasedRLEnv,
         sensor_cfg: SceneEntityCfg,
-        height_threshold: float = 0.55,
-        min_force: float = 15.0
+        height_threshold: float = 0.65,
+        min_force: float = 3.0
 ) -> torch.Tensor:
-    """带几何限制的高标准手臂支撑奖励"""
     contact_sensor = env.scene.sensors.get(sensor_cfg.name)
     if contact_sensor is None:
         return torch.zeros(env.num_envs, device=env.device)
@@ -84,10 +83,6 @@ def arm_push_up_reward(
     height_fade = torch.clamp((height_threshold - pelvis_pos_z) / 0.1, min=0.0, max=1.0)
     return total_reward * height_fade
 
-def torso_pitch_reward(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
-    proj_gravity = env.scene[asset_cfg.name].data.projected_gravity_b
-    return torch.square(proj_gravity[:, 0])
-
 def is_standing_still(
         env: ManagerBasedRLEnv,
         min_head_height: float,
@@ -123,12 +118,21 @@ def is_standing_still(
 
     return env.extras["stable_timer"] > standing_time
 
+def joint_deviation_l2(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
+    """计算关节相对于默认姿态(default_joint_pos)的偏差平方和"""
+    # 获取当前关节位置相对于默认位置的差值
+    # mdp.joint_pos_rel 返回的是 (current_pos - default_pos)
+    diff = mdp.joint_pos_rel(env, asset_cfg)
+    return torch.sum(torch.square(diff), dim=-1)
+
 # --- 2. 配置类 ---
 
 T1_JOINT_NAMES = [
     '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'
+    'Left_Ankle_Pitch', 'Right_Ankle_Pitch', 'Left_Ankle_Roll', 'Right_Ankle_Roll',
+    'AL1', 'AL2', 'AL3',
+    'AR1', 'AR2', 'AR3'
 ]
 
 
@@ -184,7 +188,7 @@ class T1ActionCfg:
 @configclass
 class T1GetUpRewardCfg:
     # 1. 姿态基础奖 (引导身体变正)
-    upright = RewTerm(func=mdp.flat_orientation_l2, weight=5.0)
+    upright = RewTerm(func=mdp.flat_orientation_l2, weight=2.0)
 
     # 2. 【条件高度奖】：双高度判定（头+盆骨），且必须脚踩地
     height_with_feet = RewTerm(
@@ -205,8 +209,8 @@ class T1GetUpRewardCfg:
         weight=15.0,  # 显著增加权重（从 3.0 提到 15.0），让它成为起步的关键
         params={
             "sensor_cfg": SceneEntityCfg("contact_sensor", body_names=[".*_hand_link", "AL3", "AR3"]),
-            "height_threshold": 0.6,  # 躯干升到 0.6m 前都鼓励手臂用力
-            "min_force": 10.0  # 只要有 15N 的力就触发
+            "height_threshold": 0.65,  # 躯干升到 0.6m 前都鼓励手臂用力
+            "min_force": 3.0  # 只要有 15N 的力就触发
         }
     )
 
@@ -217,13 +221,6 @@ class T1GetUpRewardCfg:
         params={"asset_cfg": SceneEntityCfg("robot")}
     )
 
-    # 4. 新增：躯干仰角奖 (核心诱导)
-    torso_pitch = RewTerm(
-        func=torso_pitch_reward,  # 或者使用 orientation 相关的项
-        weight=15.0,
-        params={"asset_cfg": SceneEntityCfg("robot")}
-    )
-
     # 5. 时间惩罚 (强制效率)
     time_penalty = RewTerm(
         func=mdp.is_alive,
@@ -237,6 +234,21 @@ class T1GetUpRewardCfg:
         params={"keys": "standing_success"}
     )
 
+    # 7. 手臂关节活跃度奖 (诱导摆动)
+    arm_movement_exploration = RewTerm(
+        func=mdp.joint_vel_l2,
+        weight=2,  # 权重不要太高，防止变成“风扇”
+        params={"asset_cfg": SceneEntityCfg("robot", joint_names=["AL3", "AR3"])}
+    )
+
+    # 8. 手臂位置多样性奖 (离开默认折叠姿态)
+    arm_deviation_bonus = RewTerm(
+        func=joint_deviation_l2,
+        weight=1,
+        params={
+            "asset_cfg": SceneEntityCfg("robot", joint_names=["AL3", "AR3"])
+        }
+    )
 
 @configclass
 class T1GetUpTerminationsCfg:

rl_game/get_up/env/t1_env.py

@@ -8,7 +8,7 @@ from isaaclab import sim as sim_utils
 import os
 
 _DEMO_DIR = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
-T1_USD_PATH = os.path.join(_DEMO_DIR, "asset", "t1", "t1_locomotion_physics.usd")
+T1_USD_PATH = os.path.join(_DEMO_DIR, "asset", "t1", "T1_locomotion_physics_lab.usd")
 
 @configclass
 class T1SceneCfg(InteractiveSceneCfg):