[Question] Robot control / Diffusion Policy adaptation #2
Description
Hi, thank you for the great work on Generative Modeling via Drifting!
I implemented the drifting model on simulated robot control tasks (PushT, Lift, Can, ToolHang)
by porting the drifting loss into the Diffusion Policy framework: https://github.com/phamtrongthang123/drifting_policy
The loss port passes numerical tests against your JAX implementation. Key config: gen_per_label=8, batch_size=64, dot-product cdist (matching your kernel).
However, some of my results substantially exceed the scores reported in your Table 7, which makes me think there may be a mismatch in my setup that I haven't caught. The most surprising case is ToolHang (low-dim): 0.84 vs paper's 0.38.
Full comparison:
| Task | Setting | Paper (Drifting) | My result |
|---|---|---|---|
| Can | Visual | 0.99 | 0.98 |
| PushT | Visual | 0.86 | 0.86 |
| Lift | Visual | 1.00 | 1.00 |
| PushT | State | 0.86 | 0.871 |
| Can | State | 0.98 | 0.98 |
| Lift | State | 1.00 | 1.00 |
| ToolHang | State | 0.38 | 0.84 ← large gap |
My question: Do you have plans to release an official robot control implementation?
I would not ask you to debug my code because that would be too much. But a release of official code would be very helpful.