[Question] GO1's Learned Actuator Model Evaluated at 200Hz while Trained at 50Hz

[ammousa]

Question

Hi all,

In the Go1 example, the actuator MLP is trained with 50 Hz joint data (20 ms steps), but during RL training, it’s queried at 200 Hz (50hz * decimation=4). This creates a mismatch:

• The model has never seen 5 ms transitions.
• Real actuators' dynamics were not represented at 200 Hz.

1. Shouldn’t the actuator be evaluated only every 4 sim steps (50 Hz) to match training?
2. Or should it be retrained at 200 Hz for physical consistency and better sim-to-real transfer?
3. Is there any successful deployment to a real robot using this approach in Isaaclab with a near-zero sim2real gap?

Would appreciate any insights on this design choice.

[StrainFlow]

You've got the time-step to decimation relationship backwards. If you are running the simulation with a 50 hz time step and a decimation of 4, what happens is that at each time step the physical simulation will run and actions will be applied. Every fourth step observations will be recorded and new actions will be computed.

The best way to get understand how these functions run is to debug the environment with break points at each step. This way you can see the order they run in, how often they run etc.

[ammousa]

Thanks for the clarification — I added logging to both the step() function and the actuator's compute() method for verification (see git diff below).

As shown in the attached output, the actuator MLP is evaluated once per sim step (i.e., 200 Hz), even though it was trained on 50 Hz data. This results in a 4× higher evaluation frequency than during training.

Interestingly, the actuator model is queried five times per policy step — four times within the decimation loop and once again during post-step/reset. This suggests it's being called at every sim step, not just at the action frequency as expected.

Would really appreciate any guidance on solving this issue to ensure consistency with training.

---

🔧 git diff used to instrument the print statements:

diff --git a/source/isaaclab/isaaclab/actuators/actuator_net.py b/source/isaaclab/isaaclab/actuators/actuator_net.py
index 278c2a621..d17bd35a5 100644
--- a/source/isaaclab/isaaclab/actuators/actuator_net.py
+++ b/source/isaaclab/isaaclab/actuators/actuator_net.py
@@ -175,9 +175,12 @@ class ActuatorNetMLP(DCMotor):
             )
 
         # run network inference
+        print(f"[ACTUATOR]    network input sample[0,:6]: {network_input[0,:6]}")
         with torch.inference_mode():
             torques = self.network(network_input).view(self._num_envs, self.num_joints)
+            print(f"[ACTUATOR]    network evaluated!")
         self.computed_effort = torques.view(self._num_envs, self.num_joints) * self.cfg.torque_scale
+        print(f"[ACTUATOR]    network output sample[0,:6]: {torques[0,:6]}\n")
 
         # clip the computed effort based on the motor limits
         self.applied_effort = self._clip_effort(self.computed_effort)

diff --git a/source/isaaclab/isaaclab/envs/manager_based_rl_env.py b/source/isaaclab/isaaclab/envs/manager_based_rl_env.py
index a942c99be..3e4af066a 100644
--- a/source/isaaclab/isaaclab/envs/manager_based_rl_env.py
+++ b/source/isaaclab/isaaclab/envs/manager_based_rl_env.py
@@ -182,6 +182,9 @@ class ManagerBasedRLEnv(ManagerBasedEnv, gym.Env):
         # perform physics stepping
         for _ in range(self.cfg.decimation):
             self._sim_step_counter += 1
+            print(
+                f"[STEP FUNC] Step Counter: {self.common_step_counter} | SimStep {self._sim_step_counter:03d} | Physics stepping loop counter: {_}"
+            )
             # set actions into buffers
             self.action_manager.apply_action()
             # set actions into simulator

[StrainFlow]

Can you share your environment setup? Are you using a managed or direct environment?

[ammousa]

Can you share your environment setup? Are you using a managed or direct environment?

I’m using the provided managed environment for Velocity Locomotion with the Unitree Go1. You can launch it using the following command:

./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task Isaac-Velocity-Rough-Unitree-Go1-v0 --headless

[RandomOakForest]

Thank you for following up. I'll move this post to our Discussions for the team and others to follow up. You may want to try the latest versions of the tools. Follow #3021 to reinstall. Hope this helps.