updated h1_2

legged_gym/envs/h1_2/README.md

@@ -0,0 +1,25 @@
+# H1_2 RL Example (Preview)
+
+## Simplified URDF
+
+This task utilizes a simplified version of URDF. We fix some joints and ignore most of the collisions.
+
+### Fixed Joints
+
+We fix all the joints in the hands, wrists and the elbow roll joints, since those joints have very limited effect on the whole body dynamics and are commonly controlled by other controllers.
+
+### Collisions
+
+We only keep the collision of foot roll links, knee links and the base. Early termination is majorly checked by the angular position of the base.
+
+## Dynamics
+
+Free "light" end effectors can lead to unstable simulation. Thus please be carefull with the control parameters for the joints that may affect such end effectors.
+
+## Results
+
+https://github.com/user-attachments/assets/a937e9c4-fe91-4240-88ea-d83b0160cad5
+
+## Preview Stage
+
+**The reward functions are not well tuned and cannot produce satisfactory results at the current stage. A feasible version is comming soon.**

legged_gym/envs/h1_2/h1_2_config.py

@@ -0,0 +1,235 @@
+from legged_gym.envs.base.legged_robot_config import LeggedRobotCfg, LeggedRobotCfgPPO
+
+
+class H1_2RoughCfg(LeggedRobotCfg):
+
+    class init_state(LeggedRobotCfg.init_state):
+        pos = [0.0, 0.0, 0.98]  # x,y,z [m]
+        default_joint_angles = { # = target angles [rad] when action = 0.0
+            'left_hip_yaw_joint': 0,
+            'left_hip_roll_joint': 0,
+            'left_hip_pitch_joint': -0.6,
+            'left_knee_joint': 1.2,
+            'left_ankle_pitch_joint': -0.6,
+            'left_ankle_roll_joint': 0.0,
+
+            'right_hip_yaw_joint': 0,
+            'right_hip_roll_joint': 0,
+            'right_hip_pitch_joint': -0.6,
+            'right_knee_joint': 1.2,
+            'right_ankle_pitch_joint': -0.6,
+            'right_ankle_roll_joint': 0.0,
+
+            'torso_joint': 0,
+
+            'left_shoulder_pitch_joint': 0.4,
+            'left_shoulder_roll_joint': 0,
+            'left_shoulder_yaw_joint': 0,
+            'left_elbow_pitch_joint': 0.3,
+
+            'right_shoulder_pitch_joint': 0.4,
+            'right_shoulder_roll_joint': 0,
+            'right_shoulder_yaw_joint': 0,
+            'right_elbow_pitch_joint': 0.3,
+        }
+        
+    class commands:
+        curriculum = True
+        max_curriculum = 1.
+        num_commands = 4 # default: lin_vel_x, lin_vel_y, ang_vel_yaw, heading (in heading mode ang_vel_yaw is recomputed from heading error)
+        resampling_time = 10. # time before command are changed[s]
+        heading_command = False # if true: compute ang vel command from heading error
+        class ranges:
+            lin_vel_x = [-1.0, 2.0] # min max [m/s]
+            lin_vel_y = [-1.0, 1.0]   # min max [m/s]
+            ang_vel_yaw = [-1.0, 1.0]    # min max [rad/s]
+            heading = [-3.14, 3.14]
+
+    class env(LeggedRobotCfg.env):
+        num_actions = 21
+        num_policy_output = 21
+        num_envs = 3000
+        obs_context_len = 15
+        mlp_obs_context_len = 3
+        num_observations_single = 11 + num_actions * 4 #+ 9 * 4
+        num_privileged_obs_single = 14 + num_actions * 4 #+ 9 * 4
+        num_observations = num_observations_single * obs_context_len
+        num_privileged_obs = num_privileged_obs_single * obs_context_len
+        num_teaching_observations = num_observations_single
+        
+        action_delay = 0.02
+        episode_length_s = 20
+
+    class safety:
+        # safety factors
+        pos_limit = 1.0
+        vel_limit = 1.0
+        torque_limit = 0.9
+    
+    class control(LeggedRobotCfg.control):
+        # PD Drive parameters:
+        control_type = 'P'
+        # PD Drive parameters:
+        stiffness = {
+            'hip_yaw_joint': 200.,
+            'hip_roll_joint': 200.,
+            'hip_pitch_joint': 200.,
+            'knee_joint': 300.,
+            'ankle_pitch_joint': 60.,
+            'ankle_roll_joint': 40.,
+            'torso_joint': 600.,
+            'shoulder_pitch_joint': 80.,
+            'shoulder_roll_joint': 80.,
+            'shoulder_yaw_joint': 40.,
+            'elbow_pitch_joint': 60.,
+        }  # [N*m/rad]
+        damping = {
+            'hip_yaw_joint': 5.0,
+            'hip_roll_joint': 5.0,
+            'hip_pitch_joint': 5.0,
+            'knee_joint': 7.5,
+            'ankle_pitch_joint': 1.0,
+            'ankle_roll_joint': 0.3,
+            'torso_joint': 15.0,
+            'shoulder_pitch_joint': 2.0,
+            'shoulder_roll_joint': 2.0,
+            'shoulder_yaw_joint': 1.0,
+            'elbow_pitch_joint': 1.0,
+        }  # [N*m/rad]  # [N*m*s/rad]
+        # action scale: target angle = actionScale * action + defaultAngle
+        action_scale = 0.2
+        # decimation: Number of control action updates @ sim DT per policy DT
+        decimation = 10
+
+    class asset(LeggedRobotCfg.asset):
+        file = '{LEGGED_GYM_ROOT_DIR}/resources/robots/h1_2/h1_2_simplified.urdf'
+        name = "h1_2"
+        foot_name = "ankle_roll"
+        knee_name = "knee"
+        penalize_contacts_on = ["hip", "knee"]
+        terminate_after_contacts_on = [
+            "pelvis",
+            "hip",
+            "knee",
+            "shoulder_pitch",
+            "elbow"
+        ]
+        self_collisions = 0  # 1 to disable, 0 to enable...bitwise filter
+        flip_visual_attachments = False
+        eplace_cylinder_with_capsule = False
+        collapse_fixed_joints = False
+        armature = 6e-4  # stablize semi-euler integration for end effectors
+
+        vhacd_enabled = False
+        vhacd_params_resolution = 500000
+
+    class domain_rand:
+        randomize_friction = True
+        friction_range = [0.5, 1.5]
+        randomize_base_mass = True
+        added_mass_range = [-1., 1.]
+        push_robots = True
+        push_interval_s = 15
+        max_push_vel_xy = 0.5
+        
+        randomize_gravity = False
+        gravity_range = [-1.0, 1.0]
+        gravity_rand_interval_s = 8.0 # sec
+        gravity_impulse_duration = 0.99
+        
+        dof_prop_rand_interval_s = 6
+        randomize_pd_params = True
+        kp_ratio_range = [0.5, 1.5]
+        kd_ratio_range = [0.5, 1.5]
+        randomize_motor_strength = False
+        motor_strength_range = [0.9, 1.1]
+        randomize_motor_offset = False
+        motor_offset_range = [-0.02, 0.02]
+        # randomize_decimation = False
+        # decimation_range = [0.5, 1.5]
+        soft_dof_pos_limit = 0.9
+        base_height_target = 0.98
+    
+    class sim(LeggedRobotCfg.sim):
+        dt = 0.002  # stablize semi-euler integration for end effectors
+
+    class rewards(LeggedRobotCfg.rewards):
+        soft_dof_pos_limit = 0.9
+        soft_dof_vel_limit = 0.9
+        soft_torque_limit = 0.9
+
+
+        base_height_target = 0.98
+
+        # for knee and ankle distance keeping
+        min_dist = 0.3
+        max_dist = 0.5
+
+        tracking_sigma = 0.25 # tracking reward = exp(-error^2/sigma)
+        max_contact_force = 600. # forces above this value are penalized
+        cycle_time = 0.4 #.4 #0.68
+        target_joint_pos_scale = 0.17
+        target_feet_height = 0.06
+
+        class scales(LeggedRobotCfg.rewards.scales):
+
+            joint_pos = 1.6 * 4
+            feet_clearance = 1. * 2
+            feet_contact_number = 1.2 * 1
+            feet_air_time = 1.0 * 2
+            foot_slip = -0.1
+
+            tracking_lin_vel = 1.0 * 8
+            tracking_ang_vel = 1.0 * 8
+            lin_vel_z = -0.2
+            ang_vel_xy = -0.1
+            orientation = -1
+            base_height = -100.0
+            dof_acc = -3e-8
+            collision = -1.0
+
+            action_rate = -0.6
+            action_smoothness = -0.6
+
+            dof_pos_limits = -10.0
+            dof_vel_limits = -10.0
+            torque_limits = -10.0
+            default_joint_pos = 0.5 * 4
+            feet_contact_forces = -3e-6 * 50
+            
+            ## feet reg
+            feet_parallel = -1.0
+            feet_symmetry = -1.0
+            knee_distance = 0.2
+            feet_distance = 0.2
+
+            arm_pose = -1
+            stumble = -2.0
+            stand_still = -1.0 * 0
+        
+        only_positive_rewards = True # if true negative total rewards are clipped at zero (avoids early termination problems)
+    
+    class normalization(LeggedRobotCfg.normalization):
+        clip_actions = 10.0
+
+
+class H1_2RoughCfgPPO(LeggedRobotCfgPPO):
+
+    class policy(LeggedRobotCfgPPO.policy):
+        init_noise_std = 0.3
+        activation = 'tanh'
+        policy_type = 'moving'
+        architecture = 'Trans' # choose from 'Mix', 'Trans', 'MLP', and 'RNN'
+        teaching_model_path = '/home/ziluoding/unitree_rl_gym/logs/Sep06_19-03-45_/model_13000.pt'#'/home/ps/unitree_rl_gym_o/legged_gym/model/Aug29_17-48-05_h1/model_10000.pt'
+        moving_model_path = '/home/ziluoding/humanoid-gym/logs/h1/Jul11_16-30-02_/model_12000.pt'
+
+    class runner(LeggedRobotCfgPPO.runner):
+        run_name = ''
+        experiment_name = 'h1_2'
+        run_name = ''
+        # load and resume
+        render = True
+        resume = True
+        load_run = 'Oct24_21-41-11_finetunexvel-12noheadingcycle0.40.2' # -1 = last run
+        checkpoint = -1 # -1 = last saved model
+        resume_path = None # updated from load_run and chkpt