6thSense — tactile egocentric datasets for dexterous robotics

Touch-aware demonstration data for the next generation of dexterous robots.

Robots have five senses. We are building the sixth.

Why touch-aware demonstration data

Contact onset and grip evolution are under-captured

Most datasets miss touch timing, pressure trends, and subtle adjustments that matter in dexterous manipulation.

Off-the-shelf stacks produce unusable raw dumps

Standalone sensors and recording scripts often fail calibration, synchronization, and dataset reliability requirements.

Teams need packaged data, not hardware babysitting

Robot learning teams want high-value datasets, not multi-month setup and QC work for each collection effort.

Manipulation progress is data-constrained

For contact-rich tasks, the next performance gains come from better multimodal demonstrations, not model scale alone.

How we build datasets

Hardware, sync, calibration, packaging - one stack, four stages. Robot-learning teams get aligned episodes, not raw folders.

  1. Capture — Wearable + egocentric rigs
  2. Sync — One clock, every modality
  3. Calibrate — Drift, fit, and timing checks
  4. Package — Episodes shipped model-ready

Eight aligned modalities, one episode

Tactile & pressure proxies

High-rate contact and pressure-aligned streams with per-channel calibration — where touch matters for the task.

Egocentric video

First-person RGB aligned to what the demonstrator sees — stable exposure for long household runs.

Depth (RGB-D)

Per-frame depth aligned to ego timebase for geometry, reach, and clutter around the hands.

Hand pose

Articulated hand state and grasp phases for contact-rich manipulation — not just 2D boxes in frame.

Motion & IMU dynamics

Linear acceleration, angular rates, and movement cues that characterize inertia, rhythm, and effort during the task.

Wrist & scene cameras

Secondary viewpoints for occlusion recovery, tool use, and context beyond the ego cone (roadmap / program-dependent).

Labels & dense commentary

Task and subtask boundaries, contact phases, QC flags — plus optional timestamped, frame-aligned text narration paired to video for richer training supervision.

Success / failure outcomes

Binary or graded success, failure modes, and segment-level tags for imitation and evaluation.

Representative task families

How we earn trust

Calibration boundaries, stated

We document where each signal is reliable, how drift is handled, and what should never be treated as ground-truth force.

Semantics you can train on

Pressure proxies, contact timing, and failure flags are defined so policy teams know exactly what each dimension means.

Model-ready packaging

Episodes land in the formats your trainers expect, with QC metrics and assumptions — not a dump of raw sensor folders.