Guidance Needed: How to Implement a Navigation + Grasping Task (Using Isaac-PickPlace-Locomanipulation-G1-Abs-v0 as Example)

[Sukeysun]

Question

Description
I would like to implement a navigation + grasping task, similar to the example shown in the latest branch
However, I am not sure what the recommended workflow is for building such a task in Isaac Lab.
Specifically:

Should this type of task be trained using Reinforcement Learning, Mimic Learning, or a combination of both?

What is the recommended end-to-end pipeline for navigation + manipulation tasks in Isaac Lab?

Since the repository recently added the environment
Isaac-PickPlace-Locomanipulation-G1-Abs-v0,
I would like to use this as a concrete example to understand the intended workflow.

My Questions

What is the recommended approach for navigation + grasping tasks?

Pure RL?

Pure Mimic Learning (BC / Diffusion Policy)?

Teleoperation → Mimic Learning → RL fine-tuning?

Could you provide an overview of the expected pipeline using Isaac-PickPlace-Locomanipulation-G1-Abs-v0 as the example?
For example:

Environment configuration

Teleoperation / demo collection

Mimic training

RL fine-tuning

Evaluation

If navigation + grasping is intended to follow a standard pattern in Isaac Lab, is there any documentation or reference script?

Build Info

Describe the versions that you are currently using:

• Isaac Lab Version: [e.g. 2.3.0]
• Isaac Sim Version: [5.1]

[RandomOakForest]

Thank you for posting this. I'll move this post to our Discussions for follow up. Here is a summary to consider.

For navigation + grasping tasks in Isaac Lab, particularly with the newly added Isaac-PickPlace-Locomanipulation-G1-Abs-v0 environment, you may try the following workflow: begin with mimic learning (imitation learning via teleoperated demonstrations), then optionally fine-tune with reinforcement learning (RL) for optimal performance in complex scenarios.

In summary:

• Hybrid pipeline: Start with Behavioral Cloning (BC) or Diffusion Policy using teleoperated or scripted demonstrations to quickly get a working policy. Once the policy achieves basic competency, use RL fine-tuning to improve performance, generalize, or tackle aspects hard to cover via demonstrations alone.
• Pure RL is sometimes possible, but usually sample-inefficient for long-horizon navigation + manipulation. Pure mimic learning works if expert demonstrations are high-quality, but often requires RL fine-tuning to be robust.
• The typical workflow in Isaac Lab is: Teleoperation → Mimic Learning (BC/Diffusion) → RL fine-tuning → Evaluation.

See the imitation learning documentation for reference.