second_amp

humanoid/envs/elf12/elf12_dh_stand_config.py

@@ -333,7 +333,8 @@ class Elf12DHStandCfg(LeggedRobotCfg):
         max_contact_force = 700  # forces above this value are penalized
         
         class scales:
-            ref_joint_pos = 2.2
+            # ref_joint_pos = 0
+            # ref_joint_pos = 2.2
             feet_clearance = 1.
             feet_contact_number = 2.0
             # gait
@@ -401,7 +402,7 @@ class Elf12DHStandCfgPPO(LeggedRobotCfgPPO):
     class algorithm(LeggedRobotCfgPPO.algorithm):
         entropy_coef = 0.001
         learning_rate = 1e-4
-        num_learning_epochs = 4
+        num_learning_epochs = 3
         gamma = 0.994
         lam = 0.9
         num_mini_batches = 4
@@ -425,11 +426,11 @@ class Elf12DHStandCfgPPO(LeggedRobotCfgPPO):
         load_run = -1  # -1 = last run
         checkpoint = -1  # -1 = last saved model
         resume_path = None  # updated from load_run and chkpt
-        amp_reward_coef = 1.0
+        amp_reward_coef = 0.2
         amp_motion_files = MOTION_FILES
         amp_num_preload_transitions = 2000000
         amp_task_reward_lerp = 0.9
         amp_discr_hidden_dims = [1024, 512]
         amp_replay_buffer_size = 1000000
 
-        min_normalized_std = [0.01, 0.01, 0.01, 0.01] * 3
\ No newline at end of file
+        min_normalized_std = [0.001, 0.001, 0.001, 0.001] * 3
\ No newline at end of file