Question About Diffusion Policy in Real PushT
Hi, thank you for open-sourcing this project.
I'm trying to reproduce the PushT task using SO-101.And I have a question about the real-world Push-T setting. In the simulated Push-T environment, reward can be computed directly from object/target geometry. For pusht_real_merged, how did you handle this in practice?
Did you:
- train purely with imitation learning and not use reward at all,
- use reward only for evaluation,
- estimate reward from vision / tracking,
or 4. use some manually defined success criterion?
Thanks a lot.
Hi xyin27, Thanks for your question.
We typically train with pure flow-matching/diffusion (and auxiliary) loss, which is imitation learning.
For the Evaluation, we typically observe the success rate by manual inspection of the predefined criteria.
Though you may train with an add-on with a vision-based reward model for your specific tasks (binary rewarding is better or selecting quality), when you wanna do some post-RL training.
Hope that helps.
Thanks for your reply.
I just wanted to double-check: during training, did you use only the top-camera view and the 6-dimensional state as inputs, and use only the standard Diffusion Policy implementation in LeRobot, without any additional reward model, RL fine-tuning, or major modifications?
Yes, it is doable with only imitation learning (diffusion policy), given sufficient and well-designed datasets (top-camera view and 6-DoF joint states). Recommend checking your cam fov and enabling data augmentation during training.