ParcelStow Demonstrations, Checkpoints, and Videos

Does Imitation Learning Preserve Robustness Under Execution-Timing Variation?

This repository contains 970,565 control steps from 937 successful expert demonstrations for three simulated manipulation tasks, together with learner checkpoints and evaluation records. It accompanies ParcelStow, which evaluates how the performance difference between an imitation learner and its expert varies under variation in task conditions.

Execution timing is varied through a scalar speedup factor r. It divides the nominal durations of selected phases while acquisition and settling retain fixed durations. At r=2, the scaled phases have half their nominal duration.

Support ParcelStow: ★ Star on GitHub · ♥ Like on Hugging Face

Project Page · Paper · Code

Expert–ACT Comparison

Expert on the top row; ACT below. Each column shows one task at speedup factor r=2, played at 2× speed. The ACT recordings use the checkpoints reported below. Labels identify each recorded outcome.

Open the MP4.

Nominal and Higher-Speed Task Success

Task	Expert at r=1	ACT at r=1	Higher r	Expert	ACT
Parcel insertion	100/100	100/100	2	84/100	53/100
Upright placement	185/200	194/200	2	105/200	24/200
Keyed peg insertion	182/200	191/200	1.5	150/200	3/200

Each ACT column reports one trained policy per task. Expert-minus-ACT differences at the highlighted factors are 31, 40.5, and 73.5 percentage points, with pointwise 95% paired bootstrap intervals [18,44], [32,49], and [67.5,79.5]. Parcel r=2 is within its demonstrated range; upright r=2 and peg r=1.5 are outside theirs. Training replications are reported individually in the evaluation records.

Diffusion Policy (DP) nominal success is 68/100, 181/200, and 95/200 on parcel, upright, and peg. Its parcel results support standalone success rates because episode-level pairing and distribution matching with the expert are not established. DAgger nominal success is 3/100, 1/200, and 0/200; these results support failure characterization but not execution-speed degradation claims. The record catalogs provide task-stage counts, terminal failures, hashes, pairing restrictions, and excluded series.

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	learner state observation recorded alongside the expert action
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.

Artifact Map

• data/*.parquet: viewer-friendly demonstrations, one row per 50 Hz step.
• demonstrations/*.pt: original tensor datasets used by the trainers.
• checkpoints/act_stow.pt: canonical parcel ACT-A, model seed 0.
• checkpoints/manuscript_20260921/: selected upright/peg ACT seeds 0/1/2, final DP seed-42 checkpoints, and final round-four DAgger seed-1 checkpoints.
• evaluation/manuscript_20260921/: 136 canonical conditions, 20 complete compressed source series, pairing/exclusion tables, input hashes and provenance.
• evaluation/manuscript_20260921/ARTIFACT_INDEX.csv: local code paths, HF paths, SHA-256 and sizes for all 17 canonical checkpoints and five tensor datasets.
• videos/expert_act_comparison.mp4: the expert–ACT comparison shown above.
• videos/manuscript_20260921/: individual recordings and outcome metadata.
• MANIFEST.json: checksum manifest synchronized with the code repository.

Download and Reproduce

From the code repository:

python3 scripts/download_artifacts.py --manuscript
python3 scripts/download_artifacts.py --manuscript --verify
python3 scripts/reproduce_manuscript.py --output-dir outputs/reproduce/canonical

For an offline copy, use --local-hf ../parcelstow-hf with the downloader. Existing matching tensors are skipped; corrupt existing files are not replaced. A Git LFS pointer alone is not a usable tensor file.

CPU count reproduction needs only Python; add --bootstrap and NumPy for paired intervals. It verifies original record hashes and prints 136 conditions, 20 series, 104 verified paired learner conditions and nine standalone conditions. No simulator or checkpoint is required for this operation. Original record paths are resolved through evaluation/manuscript_20260921/FILES.json.

The runbook gives explicit evaluation commands for the canonical 100-episode parcel and 200-episode upright/peg banks, ACT training and checkpoint selection, DP and DAgger training, demonstration collection, environment versions and failure checks.

Interpretation

ACT provides paired expert–learner comparisons across all three tasks. DP provides secondary task-dependent speed-response evidence. DAgger is floor-saturated at nominal speed and is retained for failure characterization. None of these results establishes hardware or simulation-to-real performance. Episode pairing and standalone restrictions are recorded per condition; do not merge overlapping records or select policies from final evaluation outcomes.

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}
}