Skip to content

Latest commit

 

History

History
51 lines (40 loc) · 3.13 KB

File metadata and controls

51 lines (40 loc) · 3.13 KB

Diffusion Policy Research Notes

Current status: closed archive. These notes document a completed baseline reproduction and several follow-up ideas that ended as negative or inconclusive results under the available compute and benchmark settings.

This folder is organized in the style of a research repository log: final verdict first, then reproduction evidence, then method attempts, then raw historical notes.

Read First

File Role Current status
00_项目收束总结.md Final project verdict Current
00_复现计划与环境记录.md Original environment and reproduction plan Historical but useful
排错日记.md Step-by-step debugging diary Evidence for reproduction
研究方向v3.6_SFC-DP与后续三步计划.md Most complete late-stage method log Final method attempt

Research Timeline

Stage File Decision
Initial planning 研究方向规划.md Broad scan of possible DP follow-ups
v2 研究方向规划v2.md Turned toward curriculum / small-data DP
v2 review 研究方向规划v2_审查与执行计划.md Narrowed execution plan after novelty/risk review
v3 研究方向v3_世界模型交叉.md Pivoted toward world-model supervision
v3.1 研究方向v3.1_预测难度重加权.md Predictive-hardness reweighting, archived as negative
v3.5 研究方向v3.5_去噪时世界模型一致性.md DLOS-DP, stopped at failed world-model gate
v3.6 review 研究方向v3.6_三方向新颖性审查与失败分析.md Consolidated failed paths and new candidates
v3.6 final 研究方向v3.6_SFC-DP与后续三步计划.md SFC-DP, inconclusive after Push-T and Lift saturation

Evidence Map

Evidence type Location
Baseline figure ../figures/pusht_baseline_seed42_ieee.png
Main experiment artifacts ../artifacts/
Analysis-only artifacts ../artifacts_analysis_only/
Predictive-hardness prototype ../pred_hardness/
DLOS prototype ../dlos_dp/
SFC-DP prototype ../sfc_dp/

Current Interpretation

The correct final interpretation is conservative:

  • Count the work as a completed Diffusion Policy baseline reproduction.
  • Preserve follow-up ideas as useful research process and negative results.
  • Do not claim a new method improvement over Diffusion Policy.
  • Resume only after changing to a more discriminative benchmark, more reliable evaluation protocol, or larger compute budget.