将一个新的呆板人模子导入最新版isaacLab举行练习（以unitree H1_2为例）

1、资产导入

• 在isaacgym中，呆板人结构文件(.urdf)生存在resources目次下
  
• 在isaaclab的文件结构里有所差别，isaaclab利用USD格式文件，此时导入新呆板人须要将urdf文件转换成USD文件。
  

1.1 文件预备

起首在宇树官方文档中拿到RL例程，把unitree_rl_gym文件夹放在根目次\home\username\下
1.2 资产导入

须要调用scripts/tools/convert_urdf.py脚本举行转换，此中涉及到一些参数的设置：
参数名形貌默认值--merge-joints布尔标志，设置为True时，归并由固定关节毗连的连杆False--fix-base布尔标志，设置为True时，将呆板人基座固定在导入位置False--joint-stiffness关节驱动的刚度，刚度值越大，关节越难变形100.0--joint-damping关节驱动的阻尼，用于镌汰关节的振动和摆动1.0--joint-target-type关节驱动的控制范例，可选值为"position"、“velocity"或"none”“position”

1. cd IsaacLab
3. ./isaaclab.sh -p scripts/tools/convert_urdf.py \
4. ~/unitree_rl_gym/resources/robots/h1_2/h1_2.urdf \
5. source/isaaclab_assets/data/Robots/h1_2/h1_2.usd \
6. --merge-joints --joint-stiffness 0.0 \
7. --joint-damping 0.0 \
8. --joint-target-type none

复制代码

导入到isaacsim中：

这个时间USD文件就天生了。
2、呆板人属性设置

对标gym中的config，在IsaacLab中也要写一个对应的config。
在IsaacLab/source/isaaclab_assets/isaaclab_assets/robots/中找到了呆板人们的设置文件，此中有一个文件为unitree.py，内里设置了Isaaclab收录的全部宇树呆板人，但是H1_2恰恰没在此中。
模拟unitree.py中关于H1的设置，再参考H1_2的关节，写一段config插在unitree.py中：
此中有几点须要注意：

• usd_path须要对应本身的usd文件的路径
  
• H1_2相比于H1而言，关节名称有些许变革（具体是名称反面多了“_joint”，以及ankle等部位多加了关节），正则表达式匹配须要在反面也加个                                    .*                              \text{.*}                  .*
  

1. H1_2_CFG = ArticulationCfg(
2.     spawn=sim_utils.UsdFileCfg(
3.         usd_path=f"/home/swanchan/IsaacLab/source/isaaclab_assets/data/Robots/h1_2/h1_2.usd",
4.         activate_contact_sensors=True,
5.         rigid_props=sim_utils.RigidBodyPropertiesCfg(
6.             disable_gravity=False,
7.             retain_accelerations=False,
8.             linear_damping=0.0,
9.             angular_damping=0.0,
10.             max_linear_velocity=1000.0,
11.             max_angular_velocity=1000.0,
12.             max_depenetration_velocity=1.0,
13.         ),
14.         articulation_props=sim_utils.ArticulationRootPropertiesCfg(
15.             enabled_self_collisions=False, solver_position_iteration_count=4, solver_velocity_iteration_count=4
16.         ),
17.     ),
18.     init_state=ArticulationCfg.InitialStateCfg(
19.         pos=(0.0, 0.0, 1.05),
20.         joint_pos={
21.             ".*_hip_yaw.*": 0.0,
22.             ".*_hip_roll.*": 0.0,
23.             ".*_hip_pitch.*": -0.16,  # -9.17 degrees
24.             ".*_knee.*": 0.36,  # 20.63 degrees
25.             ".*_ankle_pitch.*": -0.2,  # -11.46 degrees
26.             ".*_ankle_roll.*": 0.0,
27.             "torso.*": 0.0,
28.             ".*_shoulder_pitch.*": 0.4,  # 22.92 degrees
29.             ".*_shoulder_roll.*": 0.0,
30.             ".*_shoulder_yaw.*": 0.0,
31.             ".*_elbow_pitch.*": 0.3,  # 17.19 degrees
32.         },
33.         joint_vel={".*": 0.0},
34.     ),
35.     soft_joint_pos_limit_factor=0.9,
36.     actuators={
37.         "legs": ImplicitActuatorCfg(
38.             joint_names_expr=[".*_hip_yaw.*", ".*_hip_roll.*", ".*_hip_pitch.*", ".*_knee.*", "torso.*"],
39.             effort_limit=300,
40.             velocity_limit=100.0,
41.             stiffness={
42.                 ".*_hip_yaw.*": 200.0,
43.                 ".*_hip_roll.*": 200.0,
44.                 ".*_hip_pitch.*": 200.0,
45.                 ".*_knee.*": 300.0,
46.                 "torso.*": 200.0,
47.             },
48.             damping={
49.                 ".*_hip_yaw.*": 2.5,
50.                 ".*_hip_roll.*": 2.5,
51.                 ".*_hip_pitch.*": 2.5,
52.                 ".*_knee.*": 4.0,
53.                 "torso.*": 5.0,
54.             },
55.         ),
56.         "feet": ImplicitActuatorCfg(
57.             joint_names_expr=[".*_ankle_pitch.*", ".*_ankle_roll.*"],
58.             effort_limit=100,
59.             velocity_limit=100.0,
60.             stiffness={
61.                 ".*_ankle_pitch.*": 40.0,
62.                 ".*_ankle_roll.*": 40.0,
63.             },
64.             damping={
65.                 ".*_ankle_pitch.*": 2.0,
66.                 ".*_ankle_roll.*": 2.0,
67.             },
68.         ),
69.         "arms": ImplicitActuatorCfg(
70.             joint_names_expr=[".*_shoulder_pitch.*", ".*_shoulder_roll.*", ".*_shoulder_yaw.*", ".*_elbow_pitch.*"],
71.             effort_limit=300,
72.             velocity_limit=100.0,
73.             stiffness={
74.                 ".*_shoulder_pitch.*": 40.0,
75.                 ".*_shoulder_roll.*": 40.0,
76.                 ".*_shoulder_yaw.*": 40.0,
77.                 ".*_elbow_pitch.*": 40.0,
78.             },
79.             damping={
80.                 ".*_shoulder_pitch.*": 10.0,
81.                 ".*_shoulder_roll.*": 10.0,
82.                 ".*_shoulder_yaw.*": 10.0,
83.                 ".*_elbow_pitch.*": 10.0,
84.             },
85.         ),
86.     },
87. )
88. """Configuration for the Unitree H1_2 Humanoid robot."""
91. H1_2_MINIMAL_CFG = H1_2_CFG.copy()
92. H1_2_MINIMAL_CFG.spawn.usd_path = f"{ISAACLAB_NUCLEUS_DIR}/Robots/Unitree/H1_2/h1_2_minimal.usd"

复制代码

3、强化学习任务情况设置

在/IsaacLab/source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/中可以看到宇树的呆板人练习情况

直接把h1文件夹复制为h1_2，然后对内里的文件举行肯定的更改。
具体是把全部h1更换为h1_2，全部H1更换为H1_2，再改一下rough_env_cfg.py内里的关节
__init__.py

1. # Copyright (c) 2022-2025, The Isaac Lab Project Developers.
2. # All rights reserved.
3. #
4. # SPDX-License-Identifier: BSD-3-Clause
6. import gymnasium as gym
8. from . import agents
10. ##
11. # Register Gym environments.
12. ##
14. gym.register(
15.     id="Isaac-Velocity-Rough-H1_2-v0",
16.     entry_point="isaaclab.envs:ManagerBasedRLEnv",
17.     disable_env_checker=True,
18.     kwargs={
19.         "env_cfg_entry_point": f"{__name__}.rough_env_cfg:H1_2RoughEnvCfg",
20.         "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:H1_2RoughPPORunnerCfg",
21.         "skrl_cfg_entry_point": f"{agents.__name__}:skrl_rough_ppo_cfg.yaml",
22.     },
23. )
26. gym.register(
27.     id="Isaac-Velocity-Rough-H1_2-Play-v0",
28.     entry_point="isaaclab.envs:ManagerBasedRLEnv",
29.     disable_env_checker=True,
30.     kwargs={
31.         "env_cfg_entry_point": f"{__name__}.rough_env_cfg:H1_2RoughEnvCfg_PLAY",
32.         "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:H1_2RoughPPORunnerCfg",
33.         "skrl_cfg_entry_point": f"{agents.__name__}:skrl_rough_ppo_cfg.yaml",
34.     },
35. )
38. gym.register(
39.     id="Isaac-Velocity-Flat-H1_2-v0",
40.     entry_point="isaaclab.envs:ManagerBasedRLEnv",
41.     disable_env_checker=True,
42.     kwargs={
43.         "env_cfg_entry_point": f"{__name__}.flat_env_cfg:H1_2FlatEnvCfg",
44.         "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:H1_2FlatPPORunnerCfg",
45.         "skrl_cfg_entry_point": f"{agents.__name__}:skrl_flat_ppo_cfg.yaml",
46.     },
47. )
50. gym.register(
51.     id="Isaac-Velocity-Flat-H1_2-Play-v0",
52.     entry_point="isaaclab.envs:ManagerBasedRLEnv",
53.     disable_env_checker=True,
54.     kwargs={
55.         "env_cfg_entry_point": f"{__name__}.flat_env_cfg:H1_2FlatEnvCfg_PLAY",
56.         "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:H1_2FlatPPORunnerCfg",
57.         "skrl_cfg_entry_point": f"{agents.__name__}:skrl_flat_ppo_cfg.yaml",
58.     },
59. )

复制代码

flat_env_cfg.py

1. # Copyright (c) 2022-2025, The Isaac Lab Project Developers.
2. # All rights reserved.
3. #
4. # SPDX-License-Identifier: BSD-3-Clause
6. from isaaclab.utils import configclass
8. from .rough_env_cfg import H1_2RoughEnvCfg
11. @configclass
12. class H1_2FlatEnvCfg(H1_2RoughEnvCfg):
13.     def __post_init__(self):
14.         # post init of parent
15.         super().__post_init__()
17.         # change terrain to flat
18.         self.scene.terrain.terrain_type = "plane"
19.         self.scene.terrain.terrain_generator = None
20.         # no height scan
21.         self.scene.height_scanner = None
22.         self.observations.policy.height_scan = None
23.         # no terrain curriculum
24.         self.curriculum.terrain_levels = None
25.         self.rewards.feet_air_time.weight = 1.0
26.         self.rewards.feet_air_time.params["threshold"] = 0.6
29. class H1_2FlatEnvCfg_PLAY(H1_2FlatEnvCfg):
30.     def __post_init__(self) -> None:
31.         # post init of parent
32.         super().__post_init__()
34.         # make a smaller scene for play
35.         self.scene.num_envs = 50
36.         self.scene.env_spacing = 2.5
37.         # disable randomization for play
38.         self.observations.policy.enable_corruption = False
39.         # remove random pushing
40.         self.events.base_external_force_torque = None
41.         self.events.push_robot = None

复制代码

rough_env_cfg.py

1. # Copyright (c) 2022-2025, The Isaac Lab Project Developers.
2. # All rights reserved.
3. #
4. # SPDX-License-Identifier: BSD-3-Clause
6. from isaaclab.managers import RewardTermCfg as RewTerm
7. from isaaclab.managers import SceneEntityCfg
8. from isaaclab.utils import configclass
10. import isaaclab_tasks.manager_based.locomotion.velocity.mdp as mdp
11. from isaaclab_tasks.manager_based.locomotion.velocity.velocity_env_cfg import LocomotionVelocityRoughEnvCfg, RewardsCfg
13. ##
14. # Pre-defined configs
15. ##
16. from isaaclab_assets import H1_2_CFG
19. @configclass
20. class H1_2Rewards(RewardsCfg):
21.     """Reward terms for the MDP."""
23.     termination_penalty = RewTerm(func=mdp.is_terminated, weight=-200.0)
24.     lin_vel_z_l2 = None
25.     track_lin_vel_xy_exp = RewTerm(
26.         func=mdp.track_lin_vel_xy_yaw_frame_exp,
27.         weight=1.0,
28.         params={"command_name": "base_velocity", "std": 0.5},
29.     )
30.     track_ang_vel_z_exp = RewTerm(
31.         func=mdp.track_ang_vel_z_world_exp, weight=1.0, params={"command_name": "base_velocity", "std": 0.5}
32.     )
33.     feet_air_time = RewTerm(
34.         func=mdp.feet_air_time_positive_biped,
35.         weight=0.25,
36.         params={
37.             "command_name": "base_velocity",
38.             "sensor_cfg": SceneEntityCfg("contact_forces", body_names=".*ankle.*"),
39.             "threshold": 0.4,
40.         },
41.     )
42.     feet_slide = RewTerm(
43.         func=mdp.feet_slide,
44.         weight=-0.25,
45.         params={
46.             "sensor_cfg": SceneEntityCfg("contact_forces", body_names=".*ankle.*"),
47.             "asset_cfg": SceneEntityCfg("robot", body_names=".*ankle.*"),
48.         },
49.     )
50.     # Penalize ankle joint limits
51.     dof_pos_limits = RewTerm(
52.         func=mdp.joint_pos_limits, weight=-1.0, params={"asset_cfg": SceneEntityCfg("robot", joint_names=".*_ankle.*")}
53.     )
54.     # Penalize deviation from default of the joints that are not essential for locomotion
55.     joint_deviation_hip = RewTerm(
56.         func=mdp.joint_deviation_l1,
57.         weight=-0.2,
58.         params={"asset_cfg": SceneEntityCfg("robot", joint_names=[".*_hip_yaw.*", ".*_hip_roll.*"])},
59.     )
60.     joint_deviation_arms = RewTerm(
61.         func=mdp.joint_deviation_l1,
62.         weight=-0.2,
63.         params={"asset_cfg": SceneEntityCfg("robot", joint_names=[".*_shoulder_.*", ".*_elbow.*"])},
64.     )
65.     joint_deviation_torso = RewTerm(
66.         func=mdp.joint_deviation_l1, weight=-0.1, params={"asset_cfg": SceneEntityCfg("robot", joint_names="torso.*")}
67.     )
70. @configclass
71. class H1_2RoughEnvCfg(LocomotionVelocityRoughEnvCfg):
72.     rewards: H1_2Rewards = H1_2Rewards()
74.     def __post_init__(self):
75.         # post init of parent
76.         super().__post_init__()
77.         # Scene
78.         self.scene.robot = H1_2_CFG.replace(prim_path="{ENV_REGEX_NS}/Robot") # type: ignore
79.         if self.scene.height_scanner:
80.             self.scene.height_scanner.prim_path = "{ENV_REGEX_NS}/Robot/torso_link"
82.         # Randomization
83.         self.events.push_robot = None
84.         self.events.add_base_mass = None
85.         self.events.reset_robot_joints.params["position_range"] = (1.0, 1.0)
86.         self.events.base_external_force_torque.params["asset_cfg"].body_names = [".*torso_link"]
87.         self.events.reset_base.params = {
88.             "pose_range": {"x": (-0.5, 0.5), "y": (-0.5, 0.5), "yaw": (-3.14, 3.14)},
89.             "velocity_range": {
90.                 "x": (0.0, 0.0),
91.                 "y": (0.0, 0.0),
92.                 "z": (0.0, 0.0),
93.                 "roll": (0.0, 0.0),
94.                 "pitch": (0.0, 0.0),
95.                 "yaw": (0.0, 0.0),
96.             },
97.         }
99.         # Terminations
100.         self.terminations.base_contact.params["sensor_cfg"].body_names = [".*torso_link"]
102.         # Rewards
103.         self.rewards.undesired_contacts = None
104.         self.rewards.flat_orientation_l2.weight = -1.0
105.         self.rewards.dof_torques_l2.weight = 0.0
106.         self.rewards.action_rate_l2.weight = -0.005
107.         self.rewards.dof_acc_l2.weight = -1.25e-7
109.         # Commands
110.         self.commands.base_velocity.ranges.lin_vel_x = (0.0, 1.0)
111.         self.commands.base_velocity.ranges.lin_vel_y = (0.0, 0.0)
112.         self.commands.base_velocity.ranges.ang_vel_z = (-1.0, 1.0)
114.         # terminations
115.         self.terminations.base_contact.params["sensor_cfg"].body_names = ".*torso_link"
118. @configclass
119. class H1_2RoughEnvCfg_PLAY(H1_2RoughEnvCfg):
120.     def __post_init__(self):
121.         # post init of parent
122.         super().__post_init__()
124.         # make a smaller scene for play
125.         self.scene.num_envs = 50
126.         self.scene.env_spacing = 2.5
127.         self.episode_length_s = 40.0
128.         # spawn the robot randomly in the grid (instead of their terrain levels)
129.         self.scene.terrain.max_init_terrain_level = None
130.         # reduce the number of terrains to save memory
131.         if self.scene.terrain.terrain_generator is not None:
132.             self.scene.terrain.terrain_generator.num_rows = 5
133.             self.scene.terrain.terrain_generator.num_cols = 5
134.             self.scene.terrain.terrain_generator.curriculum = False
136.         self.commands.base_velocity.ranges.lin_vel_x = (1.0, 1.0)
137.         self.commands.base_velocity.ranges.lin_vel_y = (0.0, 0.0)
138.         self.commands.base_velocity.ranges.ang_vel_z = (-1.0, 1.0)
139.         self.commands.base_velocity.ranges.heading = (0.0, 0.0)
140.         # disable randomization for play
141.         self.observations.policy.enable_corruption = False
142.         # remove random pushing
143.         self.events.base_external_force_torque = None
144.         self.events.push_robot = None

复制代码

rsl_rl_ppo_cfg.py

1. # Copyright (c) 2022-2025, The Isaac Lab Project Developers.
2. # All rights reserved.
3. #
4. # SPDX-License-Identifier: BSD-3-Clause
6. from isaaclab.utils import configclass
8. from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlPpoActorCriticCfg, RslRlPpoAlgorithmCfg
11. @configclass
12. class H1_2RoughPPORunnerCfg(RslRlOnPolicyRunnerCfg):
13.     num_steps_per_env = 24
14.     max_iterations = 3000
15.     save_interval = 50
16.     experiment_name = "H1_2_rough"
17.     empirical_normalization = False
18.     policy = RslRlPpoActorCriticCfg(
19.         init_noise_std=1.0,
20.         actor_hidden_dims=[512, 256, 128],
21.         critic_hidden_dims=[512, 256, 128],
22.         activation="elu",
23.     )
24.     algorithm = RslRlPpoAlgorithmCfg(
25.         value_loss_coef=1.0,
26.         use_clipped_value_loss=True,
27.         clip_param=0.2,
28.         entropy_coef=0.01,
29.         num_learning_epochs=5,
30.         num_mini_batches=4,
31.         learning_rate=1.0e-3,
32.         schedule="adaptive",
33.         gamma=0.99,
34.         lam=0.95,
35.         desired_kl=0.01,
36.         max_grad_norm=1.0,
37.     )
40. @configclass
41. class H1_2FlatPPORunnerCfg(H1_2RoughPPORunnerCfg):
42.     def __post_init__(self):
43.         super().__post_init__()
45.         self.max_iterations = 1000
46.         self.experiment_name = "H1_2_flat"
47.         self.policy.actor_hidden_dims = [128, 128, 128]
48.         self.policy.critic_hidden_dims = [128, 128, 128]

复制代码

末了，在IsaacLab目次下实行练习脚本，就可以开始练习啦

1. ./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task Isaac-Velocity-Rough-H1_2-v0 --headless

复制代码

免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。