diff --git a/rl_game/get_up/config/ppo_cfg.yaml b/rl_game/get_up/config/ppo_cfg.yaml index 18bc16b..5c1124f 100644 --- a/rl_game/get_up/config/ppo_cfg.yaml +++ b/rl_game/get_up/config/ppo_cfg.yaml @@ -35,7 +35,7 @@ params: normalize_input: True normalize_value: True value_bootstrap: True - num_actors: 16384 # 同时训练的机器人数量 + num_actors: 8192 # 同时训练的机器人数量 reward_shaper: scale_value: 1.0 normalize_advantage: True diff --git a/rl_game/get_up/config/t1_env_cfg.py b/rl_game/get_up/config/t1_env_cfg.py index 6d63ecc..bc386a7 100644 --- a/rl_game/get_up/config/t1_env_cfg.py +++ b/rl_game/get_up/config/t1_env_cfg.py @@ -128,11 +128,15 @@ def joint_deviation_l2(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> tor # --- 2. 配置类 --- T1_JOINT_NAMES = [ + # 腿部 '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', - 'AL1', 'AL2', 'AL3', - 'AR1', 'AR2', 'AR3' + # 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' ] @@ -238,16 +242,14 @@ class T1GetUpRewardCfg: 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"])} + params={"asset_cfg": SceneEntityCfg("robot",joint_names=["Left_Shoulder.*", "Left_Elbow.*", "Right_Shoulder.*", "Right_Elbow.*"])} ) # 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"]) - } + params={"asset_cfg": SceneEntityCfg("robot",joint_names=["Left_Shoulder.*", "Left_Elbow.*", "Right_Shoulder.*", "Right_Elbow.*"])} ) @configclass @@ -272,7 +274,7 @@ class T1GetUpTerminationsCfg: @configclass class T1EnvCfg(ManagerBasedRLEnvCfg): - scene = T1SceneCfg(num_envs=16384, env_spacing=2.5) # 5090 性能全开 + scene = T1SceneCfg(num_envs=8192, env_spacing=2.5) # 5090 性能全开 def __post_init__(self): super().__post_init__()