This folder contains the v3.6 SFC-DP frequency-consistency prototype.
Archived as inconclusive. The method showed promising early signals on Push-T, but later gates did not support a robust positive claim:
- At 100 epochs, frequency supervision variants had positive signals over A.
- At 200 epochs, Push-T training saturated and the baseline caught up.
- Robomimic Lift in the tested setting saturated, with both A and E reaching
test/mean_score=1.00.
Final interpretation:
- Frequency consistency may be a useful training signal.
- The current evidence is not strong enough to claim improvement over Diffusion Policy.
- Future testing needs a harder benchmark or lower-data regime.
| File | Purpose |
|---|---|
config.py |
SFC-DP configuration |
sfc_loss.py |
Frequency-consistency loss |
train_sfc.py |
Push-T SFC-DP training script |
train_sfc_robomimic.py |
Robomimic gate training script |
stage1_generate_cmds.py |
Historical command generator |
See ../Paper/研究方向v3.6_SFC-DP与后续三步计划.md
for the full experiment log.