Robotics Data
Explorer

Humaid's integrated data explorer for browsing, inspecting, and validating multimodal robotics datasets. Review synchronized sensor streams, annotations, and metadata before training your models.

What Is the Robotics Data Explorer

The Humaid Data Explorer is a web-based interface for browsing and inspecting robotics datasets collected through human-in-the-loop data collection. It gives robotics teams direct access to the data they need to evaluate — synchronized video streams, sensor recordings, annotations, and downloadable files — without writing custom tooling or parsing raw files.

The explorer is part of the Humaid robotics data collection platform. Every dataset collected by Humaid operators — whether from teleoperation sessions, egocentric capture, or manufacturing floor recordings — is accessible through this interface.

What You Can Inspect

Dataset Catalog

Browse all available datasets organized by domain — household tasks, manufacturing assembly, and more. Each dataset includes a description, generation tag, and thumbnail previews.

Sequence Browser

Drill into individual recording sequences within each dataset. Filter and search across sequences, view metadata summaries, and select specific recordings for detailed inspection.

Synchronized Video Player

Watch dual synchronized video streams — temporal action segmentation and hand pose overlays — with frame-level playback controls. Step through individual frames or scrub to any point.

60+ Metadata Properties

Every sequence exposes detailed metadata: recording context, video specifications, annotation statistics, device information, and storage details — all visible in a structured sidebar.

Downloadable Files

Download individual data streams per sequence: raw egocentric video, third-view video, hand pose data (MP4 + JSON), 3D body pose, object detection annotations, action segmentation, and raw MCAP sensor recordings.

3D Visualization

Open sequences in the Rerun.io 3D viewer for interactive exploration of spatial data — body tracking, hand poses, and object positions rendered in three dimensions.

Data Modalities in the Explorer

The explorer surfaces the full range of multimodal data captured during Humaid's Physical AI data collection sessions. Each sequence can contain:

  • Egocentric video — first-person RGB recorded from the operator's head-mounted camera
  • Third-view video — external camera perspective for spatial context
  • 3D body pose — full-body motion tracking exported as video and NumPy arrays
  • Hand pose estimation — per-frame hand joint positions in video and JSON format
  • Object detection — bounding boxes and class labels for objects in the scene
  • Temporal action segmentation — frame-level action labels across the full recording
  • IMU and depth — inertial measurement and depth sensor data captured alongside video
  • Raw MCAP streams — the original sensor recordings in MCAP format for custom processing

How It Fits Into the Workflow

The data explorer sits at the output end of Humaid's data collection pipeline. After operators collect demonstrations, and the data passes through annotation and QA, it becomes available in the explorer for review, validation, and download.

For robotics teams, this means:

  • QA and debugging — verify annotation accuracy, check for sensor drift, confirm action labels before training
  • Training data validation — inspect individual episodes to understand what your model will learn from
  • Client delivery — browse and download delivered datasets without needing custom scripts or CLI tools
  • Team collaboration — share specific sequences and datasets with colleagues via direct URLs

Available Datasets

The explorer currently hosts open datasets across two domains:

Household Tasks

Egocentric datasets capturing everyday activities — folding, kitchen tasks, laundry, office work, room tidying, and shelf organization. Each includes 3D body tracking, hand pose estimation, object detection, and temporal action segmentation.

6 datasets · Gen 1 · Multiple recording sessions each

Manufacturing Assembly

Egocentric datasets from real production environments — component assembly, precision fitting, multi-part assembly, and quality inspection. Captured with hand-tool interaction tracking, object detection, and action segmentation.

4 datasets · Gen 2 · Production floor capture

Explore the Data

Browse real robotics datasets collected by Humaid operators. Inspect synchronized multimodal sequences, review annotations, and download the data you need.

Open Data Explorer