ParcelStow Demonstrations, Checkpoints, and Videos

This dataset accompanies ParcelStow, an Isaac Lab robot-learning benchmark for expert–learner evaluation across task execution speeds. The current arXiv v1 paper, arXiv:2609.01453, reports the parcel-insertion study.

The stable v1.0.0 software release covers parcel insertion. The GitHub main branch is active development and contains parcel insertion, upright placement, and keyed peg insertion. A v2 software release will be tagged only after the consolidated manuscript and software package are final; current main is not a released v2 package.

Task Gallery

The videos show one scripted-expert episode for each task. They illustrate individual outcomes; numerical evaluation uses the episode records in the GitHub repository.

Parcel Insertion. Acquire and reorient a parcel, then insert it into an open-front receptacle.	Upright Placement. Reorient a cuboid to an upright pose and release it in a target region.	Keyed Peg Insertion. Reorient a cuboid and insert it into a square pocket with 3 mm of clearance per side.

Three Task Configurations

Configuration	Task	Successful Expert Episodes	Demonstrated Speed Range
parcel_insertion	parcel insertion into an open-front receptacle	297	r in [0.5, 2.0]
upright_placement	cuboid reorientation to a stable upright pose in a target region	315	r in [0.75, 1.75]
peg_insertion	keyed peg insertion into a square pocket with 3 mm clearance per side	325	r in [0.5, 1.0]

Load one configuration with:

from datasets import load_dataset

task = "parcel_insertion"  # upright_placement or peg_insertion
demonstrations = load_dataset("cenwerem/parcelstow", task, split="train")

The dataset viewer exposes all three configurations through the Parquet files listed in the card metadata.

Row Schema

Each Parquet row is one 50 Hz control step:

Field	Type	Meaning
episode	int32	demonstration episode index
step	int32	control-step index within the episode
task	string	task identity
rate	float32	episode speedup factor r
observation	147 float32 values	state observation supplied to the expert
action	16 float32 values	normalized joint-position action

All three tasks share these dimensions and joint-position action semantics. The pose term is named parcel_pose for parcel insertion and object_pose for upright placement and keyed peg insertion. The normalized task-phase value also follows the selected task's own phase schedule. Task identity is therefore not an interchangeable observation semantic; see the policy interface.

Relationship among the Files

• data/*.parquet contains demonstrations in a Hugging Face Datasets layout for inspection and external training code.
• demonstrations/*.pt contains the same task demonstrations in the tensor layout read by the GitHub training scripts.
• checkpoints/*.pt contains released learned-policy checkpoints.
• GitHub data/records/ contains the frozen v1 parcel records and current development records for upright and peg. These records, not the demonstrations, reproduce the documented success counts.
• videos/ contains selected episodes and archives with JSON sidecars. Videos illustrate individual outcomes and do not replace numerical evaluation.
• GitHub artifacts/manifest.json preserves hosted paths, byte counts, and SHA-256 checksums.

Download the task-specific ACT checkpoint and .pt demonstrations without knowing manifest bundle names:

python scripts/download_artifacts.py --task parcel
python scripts/download_artifacts.py --task upright
python scripts/download_artifacts.py --task peg

Evaluation and Numerical Reproduction

Evaluate the provided scripted experts or a compatible custom policy through one GitHub command interface:

python scripts/evaluate.py --task parcel --actor expert
python scripts/evaluate.py --task upright --actor expert
python scripts/evaluate.py --task peg --actor expert

Recompute the current three-task success tables and figures without Isaac Lab or a GPU:

python3 scripts/reproduce.py all-tasks

See Reproducing the Results for record paths and the tested environment.

Checkpoint Roles and Limitations

• checkpoints/act_stow.pt is ACT-A, trained on 297 parcel demonstrations. ACT-A and the expert each succeed in 100/100 episodes at r=1, so it supplies the paper's primary nominally matched comparison.
• checkpoints/act_upright.pt is trained on 315 upright demonstrations. At r=1, ACT succeeds in 39/100 episodes and the expert in 92/100. This checkpoint is a secondary upright result, not a primary matched comparison.
• checkpoints/act_peg.pt is trained on 325 peg demonstrations. In the matched active-development records at r=1, ACT succeeds in 75/100 episodes and the expert in 93/100. At every evaluated r >= 1.5, ACT records 0/100 acquisitions. This checkpoint is a secondary peg result, not a primary matched comparison.
• Parcel insertion also provides ACT seed replicates, demonstration-count ablations, Diffusion Policy, and DAgger checkpoints. The current upright and peg development results do not provide matched Diffusion Policy or DAgger checkpoints.

All results are from Isaac Lab simulation. The dataset provides no hardware or simulation-to-real validation.

Videos

File	Playback and Content
parcel_expert_r2_2x_faster.mp4	parcel expert episode at r=2, twice the prior card playback speed
upright_expert_r1_2x_faster.mp4	upright expert episode at r=1, twice the prior card playback speed
peg_expert_r1_2x_faster.mp4	peg expert episode at r=1, twice the prior card playback speed
parcel_expert_vs_act_r2_2x_faster.mp4	parcel expert and ACT episodes at r=2; expert success and ACT orientation failure
upright_expert_vs_act_r2_2x_faster.mp4	upright expert and ACT episodes at r=2; expert success and ACT tip after release
peg_expert_vs_act_r2_2x_faster.mp4	peg expert and ACT episodes at r=2; expert success and ACT acquisition failure

Captions describe only the displayed episodes. Population success counts come from the GitHub evaluation records.

License and Citation

The dataset is licensed under Apache-2.0. Asset-specific notices remain in the GitHub repository.

@misc{enwerem2026parcelstow,
  title         = {Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds},
  author        = {Enwerem, Clinton and Baras, John S. and Belta, Calin},
  year          = {2026},
  eprint        = {2609.01453},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.01453}
}