Amend arm reward to get reward difficultly

rl_game/get_up/config/t1_env_cfg.py

@@ -52,45 +52,37 @@ def standing_with_feet_reward(
 def arm_push_up_reward(
         env: ManagerBasedRLEnv,
         sensor_cfg: SceneEntityCfg,
-        height_threshold: float = 0.6,
+        height_threshold: float = 0.55,
-        min_force: float = 2.0  # 大幅降低门槛：只要有 2N 的力就说明碰到了
+        min_force: float = 15.0
 ) -> 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)
 
-    # 1. 获取手臂受力
+    # 1. 获取受力数据
-    # 使用 net_forces_w 的范数或 Z 分量
     arm_forces_z = contact_sensor.data.net_forces_w[:, :, 2]
-    max_arm_force = torch.max(arm_forces_z, dim=-1)[0]
+    avg_arm_force = torch.mean(arm_forces_z, dim=-1)
 
-    # 2. 核心修改：将奖励分为“接触奖”和“撑地奖”
+    # 2. 几何限制：手臂必须在躯干下方 (修复了之前的 AttributeError)
-    # 接触奖：只要碰到了就给 0.2 的基础分
+    arm_body_indices, _ = env.scene["robot"].find_bodies(sensor_cfg.body_names)
-    is_contact = (max_arm_force > 0.1).float()
-
-    # 撑地奖：力在 2N 到 50N 之间线性增长 (50N 对于支撑足够了)
-    force_reward = torch.clamp((max_arm_force - min_force) / 48.0, min=0.0, max=1.0)
-
-    # 组合：有接触就有基础分，力越大加分越多
-    total_force_score = is_contact * 0.2 + force_reward * 0.8
-
-    # 3. 协同奖励：躯干高度和速度
     pelvis_idx, _ = env.scene["robot"].find_bodies("Trunk")
-    current_height = env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2]
+    pelvis_pos_z = env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2]
+    arm_pos_z = env.scene["robot"].data.body_state_w[:, arm_body_indices, 2]
+
+    # 手臂是否全部低于盆骨
+    is_below_pelvis = torch.all(arm_pos_z < pelvis_pos_z.unsqueeze(1), dim=-1).float()
+
+    # 3. 计算奖励
+    force_reward = torch.clamp((avg_arm_force - min_force) / 45.0, min=0.0, max=1.0)
     root_vel_z = env.scene["robot"].data.root_lin_vel_w[:, 2]
+    velocity_factor = torch.clamp(root_vel_z * 3.0, min=0.0, max=1.5)
 
-    # 只要在撑，且躯干在往上走，就大幅加成
+    total_reward = force_reward * is_below_pelvis * (1.0 + velocity_factor)
-    pushing_up_bonus = torch.where(
-        (max_arm_force > min_force) & (root_vel_z > 0.02),
-        total_force_score * (1.0 + torch.clamp(root_vel_z * 5.0, max=2.0)),
-        total_force_score
-    )
 
-    # 4. 这里的门槛要严格：一旦站得比较高了（比如 0.6m），就停止对手臂的奖励，
+    # 高度越高，手臂奖励越低 (强迫切换到腿)
-    # 强迫它把重心转移到腿部
+    height_fade = torch.clamp((height_threshold - pelvis_pos_z) / 0.1, min=0.0, max=1.0)
-    return torch.where(current_height < height_threshold,
+    return total_reward * height_fade
-                       pushing_up_bonus,
-                       torch.zeros_like(pushing_up_bonus))
 
 def torso_pitch_reward(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
     proj_gravity = env.scene[asset_cfg.name].data.projected_gravity_b
@@ -210,7 +202,7 @@ class T1GetUpRewardCfg:
     # 3. 手臂撑地奖：辅助脱离地面阶段
     arm_push_support = RewTerm(
         func=arm_push_up_reward,
-        weight=45.0,  # 显著增加权重（从 3.0 提到 15.0），让它成为起步的关键
+        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
@@ -235,7 +227,7 @@ class T1GetUpRewardCfg:
     # 5. 成功终极大奖
     is_success = RewTerm(
         func=lambda env, keys: env.termination_manager.get_term(keys).float(),
-        weight=50.0,
+        weight=300.0,
         params={"keys": "standing_success"}
     )