ENRIQUEMAS/PACE-Dataset-Sample

Dataset Card for PACE Dataset Sample

PACE — Procedural Action Correction Episodes — is an egocentric video dataset focused on procedural action errors, error detection, and error recovery in domestic tasks.

This repository contains a 4-episode sample of the larger proprietary PACE dataset.

Dataset Details

Dataset Description

PACE captures short first-person domestic task episodes where an intentional procedural error occurs and is followed by a structured correction or recovery pattern.

Unlike standard action-recognition datasets focused on correct task execution, PACE is designed around error-centered behavior: what happens before the mistake, when the mistake is committed, how the actor detects it, and how the task is corrected.

This sample includes 4 micro-episodes covering 4 procedural error typologies:

Episode ID	Error Type	Description
TP01_011_E1_V1	E1 — Wrong Target	Object placed in the wrong target location
TP06_002_E2_2	E2 — Wrong Object	Incorrect object used in the task
TP07_001_E3_4	E3 — Omission of Steps	Required step omitted
TP10_001_E4_7	E4 — Sequence Error	Action performed in the wrong order

• Curated by: SIANA BOX
• Shared by: ENRIQUEMAS
• Language(s) (NLP): English metadata
• License: CC-BY-NC-4.0

Dataset Sources

• Repository: Hugging Face dataset repository
• Paper: Not available
• Demo: Not available

Uses

Direct Use

This sample may be used for academic and non-commercial evaluation of:

• procedural error detection
• egocentric action recognition
• robotics task monitoring
• embodied AI research
• action segmentation
• error recovery modeling
• temporal event annotation workflows

Out-of-Scope Use

This sample is not intended for:

• commercial use without separate permission
• biometric identification
• surveillance applications
• human evaluation or profiling
• safety-critical robotics deployment without further validation
• training systems that infer personal, sensitive, or protected attributes

Dataset Structure

The downloadable sample package is:

ZIP_PACE_4_Samples.zip

ZIP_PACE_4_Samples/
  README.md
  LICENSE
  metadata/
    METADATA_PACE_4_SAMPLES_ready.csv
    METADATA_PACE_4_SAMPLES_ready.json
  videos/
    TP01_011_E1_V1.mp4
    TP06_002_E2_2.mp4
    TP07_001_E3_4.mp4
    TP10_001_E4_7.mp4

The metadata includes episode-level and temporal information related to:

task identity
episode identity
error type
error timing
PRECA / POCA structure
error detection
error correction
procedural recovery behavior
Dataset Creation
Curation Rationale

PACE was created to support research and development around procedural mistakes, not only successful task execution.

The dataset is designed around the idea that real-world robotics and embodied AI systems need to understand:

when a task is going wrong,
what kind of error occurred,
whether the error was detected,
and how recovery happens.
Source Data

The source data consists of staged egocentric videos recorded for the purpose of creating controlled procedural error episodes in household tasks.

Data Collection and Processing

Episodes were recorded as first-person domestic task videos. Each episode was selected to represent a specific error typology and then paired with structured metadata describing the procedural error and recovery pattern.


The source videos were produced by the dataset creator/team as controlled staged recordings. They are not scraped from public sources.

Annotations
Annotation Process

The episodes are annotated with temporal and procedural metadata, including the valid sequence before the error, the error action, the detection moment, the correction action, and the valid sequence after recovery.

Annotations were created by the dataset creator/team.

Personal and Sensitive Information

The dataset is focused on household object interactions and procedural actions. It is not designed to contain personal, sensitive, or protected attributes.

Users should not attempt to infer or extract personal characteristics from the videos.

Bias, Risks, and Limitations

This repository contains only 4 sample episodes and should not be treated as representative of the full commercial dataset.

Known limitations:

very small sample size
domestic-task focus only
staged procedural errors
limited number of actors/environments
not suitable for production deployment without further validation
not designed as a general-purpose human activity dataset
Recommendations

Use this sample only for evaluation, inspection, and non-commercial research testing.

For model training, benchmarking, commercial evaluation, or broader robotics use, request access to the full dataset.


Glossary
PACE: Procedural Action Correction Episodes
PRECA: Valid procedural action sequence before the error
POCA: Valid procedural action sequence after error
Error Action: The physical moment where the procedural mistake occurs
Error Detection: The moment where the person appears to realize the mistake
Error Correction: The physical recovery action used to restore the task
E1 — Wrong Target: Correct object placed in the wrong target location
E2 — Wrong Object: Incorrect object used
E3 — Omission of Steps: Required action omitted
E4 — Sequence Error: Action performed in the wrong order
More Information

The full commercial PACE dataset contains 1,400+ annotated episodes across multiple household task typologies and controlled detection/correction variants.

To request full access, commercial licensing, or academic collaboration, contact:


PACE Dataset Access
Dataset Card Authors

SIANA BOX / ENRIQUEMAS

Dataset Card Contact

em@sianabox.com