The example of train-T1

rl_game/demo/config/ppo_cfg.yaml

@@ -0,0 +1,60 @@
+params:
+  seed: 42
+  algo:
+    name: a2c_continuous
+
+  model:
+    name: continuous_a2c_logstd
+
+  network:
+    name: actor_critic
+    separate: False
+    space:
+      continuous:
+        mu_activation: None
+        sigma_activation: None
+        mu_init:
+          name: default
+        sigma_init:
+          name: const_initializer
+          val: 0
+        fixed_sigma: True
+    mlp:
+      units: [512, 256, 128]
+      activation: elu
+      d2rl: False
+      initializer:
+        name: default
+
+  config:
+    name: T1_Walking
+    env_name: rlgym # Isaac Lab 包装器
+    multi_gpu: False
+    ppo: True
+    mixed_precision: True
+    normalize_input: True
+    normalize_value: True
+    value_bootstrap: True
+    num_actors: 16384 # 同时训练的机器人数量
+    reward_shaper:
+      scale_value: 1.0
+    normalize_advantage: True
+    gamma: 0.99
+    tau: 0.95
+    learning_rate: 3e-4
+    lr_schedule: adaptive
+    kl_threshold: 0.008
+    score_to_win: 20000
+    max_epochs: 5000
+    save_best_after: 50
+    save_frequency: 100
+    grad_norm: 1.0
+    entropy_coef: 0.01
+    truncate_grads: True
+    bounds_loss_coef: 0.0
+    e_clip: 0.2
+    horizon_length: 128
+    minibatch_size: 32768
+    mini_epochs: 5
+    critic_coef: 2
+    clip_value: True

rl_game/demo/config/t1_env_cfg.py

@@ -0,0 +1,94 @@
+from isaaclab.envs import ManagerBasedRLEnvCfg
+from isaaclab.managers import ObservationGroupCfg as ObsGroup
+from isaaclab.managers import ObservationTermCfg as ObsTerm
+from isaaclab.managers import RewardTermCfg as RewTerm
+from isaaclab.managers import TerminationTermCfg as DoneTerm
+from isaaclab.envs.mdp import JointPositionActionCfg
+import isaaclab.envs.mdp as mdp
+from isaaclab.utils import configclass
+from rl_game.demo.env.t1_env import T1SceneCfg
+
+@configclass
+class T1ObservationCfg:
+    """观察值空间配置容器"""
+
+    @configclass
+    class PolicyCfg(ObsGroup):
+        concatenate_terms = True
+        enable_corruption = False
+
+        # ⬅️ 2. 修改点：直接使用 mdp.函数名，不要引号
+        base_lin_vel = ObsTerm(func=mdp.base_lin_vel)
+        base_ang_vel = ObsTerm(func=mdp.base_ang_vel)
+        projected_gravity = ObsTerm(func=mdp.projected_gravity)
+        joint_pos = ObsTerm(func=mdp.joint_pos_rel)
+        joint_vel = ObsTerm(func=mdp.joint_vel_rel)
+        actions = ObsTerm(func=mdp.last_action)
+
+    policy = PolicyCfg()
+
+@configclass
+class T1ActionCfg:
+    """动作空间配置"""
+    joint_pos = JointPositionActionCfg(
+        asset_name="robot", # 注意这里是 asset_name，对应场景里的机器人名称
+        joint_names=[".*"], # 控制所有关节
+        scale=0.5, # 缩放网络输出
+        use_default_offset=True # 动作是相对于默认关节角度(init_state里的0)的偏移
+    )
+
+@configclass
+class T1TerminationsCfg:
+    """终止条件：什么时候重置环境"""
+    # 1. 摔倒重置：如果躯干高度低于 0.35米 (假设 T1 胯部在 0.7米)
+    base_height_too_low = DoneTerm(
+        func=mdp.root_height_below_minimum,
+        params={"minimum_height": 0.35},
+    )
+    # 2. 存活时间限制 (Timeout)
+    time_out = DoneTerm(func=mdp.time_out)
+
+@configclass
+class T1CommandsCfg:
+    """命令配置：定义机器人的目标速度"""
+    base_velocity = mdp.UniformVelocityCommandCfg(
+        asset_name="robot",
+        resampling_time_range=(10.0, 10.0),
+        ranges=mdp.UniformVelocityCommandCfg.Ranges(
+            lin_vel_x=(0.5, 1.5),
+            lin_vel_y=(0.0, 0.0),
+            ang_vel_z=(-0.1, 0.1),
+        ),
+    )
+
+@configclass
+class T1RewardCfg:
+    """奖励函数配置：鼓励向前走，惩罚摔倒和过大能耗"""
+    # 速度追踪奖励 (假设目标是沿 X 轴走)
+    track_lin_vel_xy_exp = RewTerm(
+        func=mdp.track_lin_vel_xy_exp,
+        weight=1.0,
+        params={
+            "std": 0.5,
+            "command_name": "base_velocity"
+        }
+    )
+    # 姿态惩罚 (保持上半身直立)
+    upright = RewTerm(func=mdp.flat_orientation_l2, weight=0.1)
+    # 动作平滑惩罚
+    action_rate_l2 = RewTerm(func=mdp.action_rate_l2, weight=-0.01)
+
+@configclass
+class T1EnvCfg(ManagerBasedRLEnvCfg):
+    """主环境配置"""
+    # 场景设置
+    scene = T1SceneCfg(num_envs=16384, env_spacing=2.5)
+    # 观察与奖励
+    observations = T1ObservationCfg()
+    rewards = T1RewardCfg()
+    terminations = T1TerminationsCfg()
+    actions = T1ActionCfg()
+    commands = T1CommandsCfg()
+    episode_length_s = 20.0
+    # 默认步长
+    decimation = 6 # 仿真频率/控制频率