erdrsvhr9/temporally_conditioned_diffusion_policy

TC-DP paper reproduction data

Demonstrations, selected checkpoints and evaluation rollouts for the main three-training-seed comparison in Mitigating Temporal Failures in Diffusion-Based Robotic Imitation Learning. Rollout images are retained as uint8 pixels.

Current paper selection

• demonstration/: WaitAtGoal, LiftQA and LiftQA-Real demonstrations.
• checkpoints/: 21 selected DP/TC-DP checkpoints and their configurations.
• inference/: 28 evaluation datasets, including the separate PIP evaluations.
• ldp/outputs/: six selected PTP checkpoints and resolved configurations.
• ldp/inference/: six PTP evaluations (three seeds for each simulated task).

Zarr datasets are distributed as .zarr.zip archives. Each archive contains the original .zarr directory, all its arrays (including images), and metadata. Extract beside the archive to restore the dataset. PTP uses the same WaitAtGoal and LiftQA demonstrations; these are not duplicated.

Download and reproduce

In the matching TemporalDiffusionPolicy code release, run:

python src/temporal_diffusion_policy/tools/dataset_upload.py --download
python -m temporal_diffusion_policy.analysis.report

In the matching ldp release, download the PTP files into data/ with:

python dataset_upload.py --download

Download the shared demonstrations with TemporalDiffusionPolicy. Set paths in analysis study lists and training configurations to your local repository paths; PTP studies should point at the downloaded ldp data/inference/ directories. The default paper_main report uses three training seeds for WaitAtGoal and LiftQA and the single-policy LiftQA-Real evaluations. It reports means and sample standard deviations across training seeds for the simulated tasks.

The matching TemporalDiffusionPolicy code's tcdp_paper/paper_upload_files.json and ldp's paper_upload_files.json list the exact current-paper selection. The upload adds these files without deleting older hosted experiments. The selection excludes ablations and other exploratory runs.

Robomimic checkpoints and rollouts are hosted separately at https://huggingface.co/datasets/erdrsvhr9/temporally_conditioned_diffusion_policy_robomimic. Can, Lift and Square demonstrations should be obtained from robomimic upstream.

Citation

Temporary citation for this repository during anonymous review. Replace this entry with the final publication details when available.

@misc{tcdp2026,
  title={Mitigating Temporal Failures in Diffusion-Based Robotic Imitation Learning},
  author={{Anonymous Authors}},
  year={2026},
  howpublished={Hugging Face dataset repository},
  url={https://huggingface.co/datasets/erdrsvhr9/temporally_conditioned_diffusion_policy},
  note={Anonymous review version; temporary citation}
}