add joint_pos_penalty_l1 for better performance.

source/robot_lab/robot_lab/tasks/go2/env_cfg.py

@@ -366,7 +366,7 @@ class RewardsCfg:
     # Thus, we need to use a smaller weight for the dof_acc_l2 term in Lab compared to Gym.
     dof_acc_l2 = RewTerm(
         func=mdp.joint_acc_l2, 
-        weight=-1.0e-7, # gym和lab的dof_acc reward实现尺度不一样，lab是physic step level的，由于l2对于离群值的敏感性会导致更大的惩罚
+        weight=-1.0e-7,
         params={"asset_cfg": SceneEntityCfg("robot", joint_names=JOINT_NAMES)}
     )
     joint_power = RewTerm(
@@ -419,6 +419,17 @@ class RewardsCfg:
             "command_threshold": 0.1,
         },
     )
+    joint_pos_penalty_l1 = RewTerm(
+        func=mdp.joint_pos_penalty_l1,
+        weight=-0.02,
+        params={
+            "command_name": "base_velocity",
+            "asset_cfg": SceneEntityCfg("robot", joint_names=".*_(thigh|calf)_joint"),
+            "stand_still_scale": 1.0,
+            "velocity_threshold": 0.1,
+            "command_threshold": 0.1,
+        },
+    )    
     
 @configclass
 class TerminationsCfg:

source/robot_lab/robot_lab/tasks/go2/mdp/rewards.py

@@ -109,7 +109,7 @@ def stand_still(
     return reward
 
 
-def joint_pos_penalty(
+def joint_pos_penalty_l1(
     env: ManagerBasedRLEnv,
     command_name: str,
     asset_cfg: SceneEntityCfg,
@@ -123,14 +123,13 @@ def joint_pos_penalty(
     cmd = torch.linalg.norm(env.command_manager.get_command(command_name), dim=1)
     body_vel = torch.linalg.norm(asset.data.root_lin_vel_b[:, :2], dim=1)
     running_reward = torch.linalg.norm(
-        (asset.data.joint_pos[:, asset_cfg.joint_ids] - asset.data.default_joint_pos[:, asset_cfg.joint_ids]), dim=1
+        (asset.data.joint_pos[:, asset_cfg.joint_ids] - asset.data.default_joint_pos[:, asset_cfg.joint_ids]), dim=1, ord=1
     )
     reward = torch.where(
         torch.logical_or(cmd > command_threshold, body_vel > velocity_threshold),
         running_reward,
         stand_still_scale * running_reward,
     )
-    reward *= torch.clamp(-env.scene["robot"].data.projected_gravity_b[:, 2], 0, 0.7) / 0.7
     return reward