Anonymous-HEIR123/HEIR-Dataset

HEIR: Harness Egocentric Intent for Human–Robot Interactions

This release contains 300 processed session annotation JSON files from EgoHRI/HEIR, organized into 240 training sessions and 60 test sessions. The 22 additional reviewed sessions are included in training.

Project website · Download all 300 annotations

Release contents

Split	Sessions	Episodes	Execution subtasks
Training	240	1,466	5,084
Test	60	370	1,332
Total	300	1,836	6,416

The subtasks comprise 2,744 navigation and 3,672 manipulation segments. The split is defined at the session level; it is not a participant-disjoint split.

• heir_annotations.zip: exactly 300 JSON annotations under train/ and test/.
• manifest.csv: session IDs, split assignments, paper-aligned task names and stable task IDs, episode/subtask counts, archive paths, and SHA-256 checksums.
• task_name_mapping.json: all 20 task names from Figure 3, with stable task IDs, domains, and session counts.
• annotation_schema.json: JSON Schema for the released session records, including optional instruction events.
• release_summary.json: counts and processing/validation summary.
• verify_release.py: standalone archive and structure verification using the Python standard library.
• SHA256SUMS: checksums for the release files.

This is an annotation-only release. It does not contain videos, images, audio, gaze streams, robot trajectories, Parquet media data, reconstruction results, model weights, or intermediate annotation versions. Media paths inside the annotations are relative references to files that are not included in this repository.

Task names

The release contains 20 task categories, with names aligned to Figure 3 of the paper. Every category's session count matches the annotations. Home, Work, and Kitchen contain 110, 62, and 128 sessions, respectively.

Examples: room_bed → Bed making; work_supply → Office supplies; kitchen_fruit → Afternoon tea serving. The full mapping is in task_name_mapping.json.

Annotation structure

Each file describes one continuous interaction session:

session
  session_id, split, task_id, task_name, task_domain, participant_ids
  relative human/robot media references and alignment metadata
  episodes[]
    episode_id and time/frame boundaries
    instruction: original utterance and English translation
    intent_label: grounded intent, manipulated_object[], goal_entity[], full_task
    robot_execution
      overall time/frame/Parquet boundaries and outcome
      subtasks[]
        subtask_id, subtask_type
        time/frame/Parquet boundaries
        subgoal, completion_condition
    instruction_events[] (optional)

manipulated_object and goal_entity are lists despite their singular field names. Subtasks are located inside each episode's robot_execution. Time values are in seconds; frame and Parquet indices retain the original annotation conventions. Temporal overlaps between episodes are preserved and must not be interpreted as independently recorded sessions. Hierarchical episode/subtask IDs are scoped to their session.

Most semantic labels and translations are in English. instruction.text_raw preserves the original Chinese utterance, with personal names pseudonymized where present. Optional instruction events preserve references to related subtasks and timing evidence metadata; evidence source paths have been removed and evidence media are not included.

Processing and validation

• Replaced recording identifiers containing names/dates with opaque release IDs (HEIR_S0001–HEIR_S0300).
• Replaced participant names with consistent pseudonyms (P01–P17) and removed collection identifiers and private audit paths.
• Replaced explicit personal-name mentions in 5 text fields with consistent pseudonyms. Numeric values, boundaries, outcomes, relative media references, and all other semantic annotations are preserved.
• Aligned every session-level task_name with Figure 3's Session Design labels. task_id preserves the original machine-readable category identifier; task_domain uses Home, Work, or Kitchen. This does not rewrite intent descriptions, plans, or execution subgoals.
• Preserved optional instruction-event records and all references between episodes and subtasks.
• Kept the original-to-release mapping private; it is not part of the public repository.
• Verified JSON/schema validity, manifest checksums, session/episode/subtask counts, unique hierarchical IDs, interval ordering, object/entity list counts, and instruction-event references.

Participant pseudonyms originate from directory-alias normalization; they have not been independently verified against a participant registry. This package has not undergone a fresh video-by-video semantic reannotation. The checks above validate the exported structure and processing, not every underlying label's correctness.

Load and verify

Download the archive and accompanying files from Files and versions, then run:

python3 verify_release.py

Read one annotation without extracting the archive:

import json
import zipfile

with zipfile.ZipFile("heir_annotations.zip") as archive:
    train_files = sorted(name for name in archive.namelist() if name.startswith("train/"))
    session = json.loads(archive.read(train_files[0]))
    print(session["session_id"], len(session["episodes"]))

The archive contains the actual annotations; manifest.csv is an index rather than an additional set of annotations. The Hugging Face Dataset Viewer displays this 300-row session index, not the nested annotation JSON files. Downloading annotations alone does not provide the visual observations or action trajectories needed to reproduce every multimodal benchmark.

License

This repository retains its existing CC BY 4.0 license metadata for the released annotations. This annotation release does not grant access to the underlying recordings.