EgoExo-Wreck: Synchronized Egocentric/Exocentric Video — Object Destruction, Debris Manipulation & Cleanup

Data Product Features

• Paired video streams: head-mounted egocentric 4K (GoPro Hero 12, 3840×2160 @ 30/60 fps) + fixed-mount exocentric (2688×1520 or 1920×1080 @ 30 fps), per session
• Per-clip synchronization data: ego/exo sync frame anchors, lag in seconds, confidence rating, and method (audio waveform correlation, 8 kHz mono RMS-delta envelope; ±~50 ms accuracy)
• Machine-readable manifest (CSV): clip IDs, session IDs, capture dates, activity class, room/location, durations, synchronized-overlap runtime, device metadata, frame rates, file sizes
• Object-detection auto-labels: 16-class in-domain detector (room object inventory including tools), with documented training methodology and per-class metrics
• Narration layer (newer cleanup capture): staff natural-language narration recorded on a separate audio channel with timestamped transcripts
• Multi-person sessions: 1–6 participants per destruction session, including multi-agent coordination footage

Distribution

• Format: MP4 video (H.264/H.265); CSV manifest; label files (YOLO/COCO format); narration transcripts (text with timestamps)
• Data Volume: ~150 hours of synchronized ego-exo video across paired clips (per-clip synchronized-overlap runtime documented in manifest); growing ~80 hours/month; typical clip length 6–31 minutes; ego files ~1.5–20 GB per clip, exo files ~0.4–1 GB per clip
• Structure: session-based pairing — each record links one egocentric file to its matched exocentric file with frame-level sync anchors
• Delivery: secure file transfer or cloud bucket; evaluation sample available under click-through agreement

Usage

• Robot manipulation / imitation learning: demonstrations of grasping, sweeping, and handling irregular, deformable, and fragmented objects (broken glass, ceramic shards, mixed debris) in unstructured settings
• Vision-language-action (VLA) model training: paired ego-exo demonstrations with natural-language narration for instruction-conditioned learning
• World models / video generation: real impact dynamics, shattering, and object state transitions (intact → damaged → fragment) from synchronized viewpoints
• Cross-view understanding: ego-exo correspondence, view-invariant action recognition, cross-view retrieval
• Activity recognition and temporal segmentation: repeated procedural cycles with consistent scenes and natural variation
• Physics and damage modeling: before/during/after sequences of object destruction for simulation ground truth and damage-assessment models

Coverage

• Geographic Coverage: United States (single facility, Las Vegas, Nevada) — 4 destruction rooms plus warehouse/storage reset areas
• Time Range: January 2026 – present (ongoing monthly capture)
• Demographics: adult participants, 1–6 per session; staff-performed cleanup sequences; demographic attributes not annotated

License

AI Training Rights

• Use the Data Product to train, fine-tune, and evaluate machine learning models, including large language models.
• Incorporate Data Product content into models and commercialize resulting model outputs.
• Create derivative works (model weights, embeddings, etc.) for any lawful purpose.

• The Data Product itself may not be sold, redistributed, or shared outside of licensed usage.
• Licensee must comply with all applicable laws, including data protection and privacy regulations.

Who Can Use It

• Robotics and embodied-AI labs: training manipulation policies and VLA models on human demonstrations of debris handling and object interaction
• World-model and video-generation teams: learning physical dynamics from real destruction and state-transition footage
• Computer vision researchers: ego-exo benchmarks, cross-view correspondence, action recognition, temporal segmentation
• Simulation, game, and VFX developers: real shatter and impact reference for physics simulation ground truth
• Insurance and damage-assessment AI developers: before/after damage sequences for assessment-model training

Data Dictionary