[Question] How to define one RL “step” as a full grasp execution (not physics dt) in Isaac Lab?

[qkrwnduf1997]

Question

Hi Isaac Lab team,

I’m working on 4-DoF robotic grasp pose detection from RGB-D images. The setup is top-down RGB-D sensing and gripper&manipulator to grasp and relocate objects.

Observation: single RGB-D image (top-down)
Action: grasp pose (x,y,z,yaw)
Reward: binary success (1 if the executed grasp succeeds, 0 otherwise)

Desired step semantics: one RL step = one complete grasp attempt (move → close → lift → success/fail), not the physics timestep dt.

From what I can tell, Isaac Lab’s stepping appears tied to the physics timestep, and most examples are continuous control (joint velocity/torque) with a reward each dt. In my case, I need to aggregate multiple simulation timesteps into a single RL step and return one reward only after the grasp attempt completes.

Questions

• Is there an official/recommended way in Isaac Lab to define a higher-level RL step that bundles many simulation timesteps into one environment step (i.e., event-driven or option/macro-action style stepping)?

• Are there reference examples where a single agent step triggers a closed-loop sequence over several physics steps (e.g., pick-and-place or grasp attempts) and returns a single terminal/non-terminal reward at the end of that sequence?

• If the intended approach is to run internal controllers over multiple substeps, what’s the recommended pattern (e.g., custom pre_physics_step / post_physics_step with an internal state machine, step-skipping/frame-skip, or a callback to signal “grasp-attempt complete” before exposing step() to the agent)?

[RandomOakForest]

Thank you for posting these questions. I'll move the post to our Discussions section for follow up. In general, you'l'l have to implement custom environment logic: Place code in environment methods like pre_physics_step, post_physics_step, or similar hooks to execute the multi-timestep grasp logic.