trainable

Makefile

@@ -33,7 +33,7 @@ train-many-vis:
 	@python -m sim.humanoid_gym.train --task humanoid_ppo --run_name v1 --num_envs 16
 
 train:
-	@python -m sim.humanoid_gym.train --task humanoid_ppo --run_name v1 --num_envs 2048 --headless
+	@python -m sim.humanoid_gym.train --task humanoid_ppo --run_name v1 --num_envs 6144 --headless
 
 play:
 	@python -m sim.humanoid_gym.play --task humanoid_ppo --run_name v1

sim/humanoid_gym/envs/humanoid_config.py

@@ -82,18 +82,20 @@ class init_state(LeggedRobotCfg.init_state):  # noqa: N801
 
     class sim(LeggedRobotCfg.sim):  # noqa: N801
         dt = 0.005
-        substeps = 1
+        substeps = 2
 
         class physx(LeggedRobotCfg.sim.physx):  # noqa: N801
             num_threads = 10
             solver_type = 1
             num_position_iterations = 4
             num_velocity_iterations = 1
-            contact_offset = -0.01
+            # contact_offset = -0.01
+            # rest_offset = -0.015
+            contact_offset = 0.01
             rest_offset = -0.015
             bounce_threshold_velocity = 0.5
-            max_depenetration_velocity = 10.0
-            max_gpu_contact_pairs = 2**23
+            max_depenetration_velocity = 1.0
+            max_gpu_contact_pairs = 2**24
             default_buffer_size_multiplier = 5
             contact_collection = 2  # gymapi.CC_ALL_SUBSTEPS
 
@@ -157,14 +159,10 @@ class scales:  # noqa: N801
             base_acc = 0.0
 
             # Energy
-            # torques = -1e-6
-            # dof_vel = -5e-5
-            # dof_acc = -1e-8
-            # collision = -1e-2
-            torques = 0.0
-            dof_vel = 0.0
-            dof_acc = 0.0
-            collision = 0.0
+            torques = -1e-6
+            dof_vel = -5e-5
+            dof_acc = -1e-8
+            collision = -1e-2
 
     class normalization:  # noqa: N801
         class obs_scales:  # noqa: N801

sim/humanoid_gym/envs/humanoid_env.py

@@ -211,7 +211,8 @@ def reset_idx(self, env_ids: Tensor) -> None:
     def _compute_torques(self, actions: Tensor) -> Tensor:
         # Need to override this so we can just use the motor torques directly.
         # return torch.clip(actions * self.cfg.control.action_scale, -self.torque_limits, self.torque_limits)
-        return torch.tanh(actions * self.cfg.control.action_scale) * self.torque_limits
+        torques = torch.tanh(actions * self.cfg.control.action_scale) * self.torque_limits
+        return torques
 
     def check_termination(self) -> None:
         self.reset_buf = self.episode_length_buf > self.max_episode_length