The demo of get up

rl_game/get_up/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: relu
+      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: 0.96
+    tau: 0.95
+    learning_rate: 5e-4
+    lr_schedule: adaptive
+    kl_threshold: 0.013
+    score_to_win: 20000
+    max_epochs: 5000
+    save_best_after: 50
+    save_frequency: 100
+    grad_norm: 1.0
+    entropy_coef: 0.02
+    truncate_grads: True
+    bounds_loss_coef: 0.001
+    e_clip: 0.2
+    horizon_length: 32
+    minibatch_size: 16384
+    mini_epochs: 4
+    critic_coef: 2
+    clip_value: True

rl_game/get_up/config/t1_env_cfg.py

@@ -0,0 +1,116 @@
+from isaaclab.assets import ArticulationCfg
+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.managers import EventTermCfg as EventTerm
+from isaaclab.envs.mdp import JointPositionActionCfg
+from isaaclab.managers import SceneEntityCfg
+from isaaclab.utils import configclass
+from rl_game.get_up.env.t1_env import T1SceneCfg
+
+import isaaclab.envs.mdp as mdp
+
+
+@configclass
+class T1ObservationCfg:
+    """观察值：必须包含姿态和高度，否则机器人不知道自己是怎么躺着的"""
+
+    @configclass
+    class PolicyCfg(ObsGroup):
+        concatenate_terms = True
+        enable_corruption = False
+
+        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)
+        # ⬅️ 新增：让机器人知道自己离地多高
+        root_height = ObsTerm(func=mdp.root_height)
+        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 T1EventCfg:
+    """事件配置：实现多种随机躺下姿态的关键"""
+
+    # 每次重置时，随机给基座一个旋转角度
+    # 包括：背躺 (Pitch=1.57), 趴着 (Pitch=-1.57), 侧躺 (Roll=1.57)
+    reset_robot_rotation = EventTerm(
+        func=mdp.reset_root_custom_orientation,
+        params={
+            "euler_angles_range": {
+                "roll": (-1.57, 1.57),  # 覆盖侧躺
+                "pitch": (-1.57, 1.57),  # 覆盖趴下和背躺
+                "yaw": (-3.14, 3.14),  # 随机朝向
+            },
+            "asset_cfg": SceneEntityCfg("robot"),
+        },
+        mode="reset",
+    )
+
+
+@configclass
+class T1GetUpRewardCfg:
+    """起身任务专用奖励：快、直、稳"""
+
+    # 1. 核心进度：高度奖励
+    # 只要低于目标高度(0.6m)，就会持续根据距离给负分，越高分数越接近0
+    height_progress = RewTerm(
+        func=mdp.root_height_below_minimum,
+        weight=15.0,
+        params={"minimum_height": 0.65}
+    )
+
+    # 2. 速度惩罚：时间就是金钱
+    # 每一帧都扣 0.5 分，起身越慢，被扣掉的总分就越多
+    time_penalty = RewTerm(
+        func=mdp.is_alive,
+        weight=-0.5
+    )
+
+    # 3. 姿态奖励：上半身必须垂直向上
+    upright = RewTerm(func=mdp.flat_orientation_l2, weight=2.0)
+
+    # 4. 平滑惩罚：防止起身时由于急躁导致的剧烈抖动
+    action_rate = RewTerm(func=mdp.action_rate_l2, weight=-0.01)
+    joint_vel = RewTerm(func=mdp.joint_vel_l2, weight=-0.001)
+
+
+@configclass
+class T1TerminationsCfg:
+    """终止条件"""
+    # 提前成功：如果站起来了并保持稳定，直接结束 Episode 并拿奖励，节省算力
+    is_standing_success = DoneTerm(
+        func=mdp.root_height_below_minimum,
+        params={"minimum_height": 0.65, "inverted": True}  # 高于 0.65m 则重置
+    )
+    time_out = DoneTerm(func=mdp.time_out)
+
+
+@configclass
+class T1EnvCfg(ManagerBasedRLEnvCfg):
+    """主环境配置：针对 T1 起身任务"""
+    scene = T1SceneCfg(num_envs=16384, env_spacing=2.5)
+
+    # 覆盖初始位置：必须低高度生成，防止随机旋转后高空坠落
+    def __post_init__(self):
+        super().__post_init__()
+        self.scene.robot.init_state.pos = (0.0, 0.0, 0.2)
+
+    observations = T1ObservationCfg()
+    rewards = T1GetUpRewardCfg()
+    terminations = T1TerminationsCfg()
+    events = T1EventCfg()  # ⬅️ 挂载随机姿态事件
+
+    actions = JointPositionActionCfg(
+        asset_name="robot", joint_names=[".*"], scale=0.5, use_default_offset=True
+    )
+
+    # 训练参数
+    episode_length_s = 5.0  # 强制 5秒内必须站起来
+    decimation = 4  # 提高控制频率 (约 50Hz) 有助于精细动作