Robot Data Atlas

Every signal needed to train robot foundation models.

Modern robot models — GR00T, π0/π0.5, OpenVLA, Helix, Octo — train on a heterogeneous mix of real teleop, autonomous rollouts, simulation, human video, motion capture, force/tactile, and language data. This atlas maps each class to the public datasets that pioneered it and the SignIQ capability that supplies it commercially.

Real Robot Action-State Trajectories

Available now

Synchronized observations, proprioception, action vectors, gripper state, and language instructions captured directly from real robots performing manipulation tasks.

Sample
4 live · interactive
Export
LeRobot v2/v2.1 · GR00T-LeRobot · …

Signals

RGBDepthProprioceptionActionGripperLanguageSuccess/Failure

Used for

PretrainingPost-trainingFine-tuningEvaluation

Models that train on this

GR00T · openpi (π0/π0.5) · OpenVLA · Octo · ACT · Diffusion Policy

See it in our Sample Lab

SignIQ capability

Calibrated multi-camera capture with frame-accurate timestamps, master/puppet teleop streams, and per-episode QA before delivery.

Autonomous Policy Rollouts

Available now

Trajectories generated by deployed policies, including success and failure cases. Critical for post-training, recovery skill learning, and reliability evaluation.

Sample
1 live · interactive
Export
LeRobot v2/v2.1 · HDF5 · …

Signals

RGBActionStateOutcome labelFailure-mode tag

Used for

Post-trainingEvaluation

Models that train on this

GR00T fine-tuning · Skill recovery policies · VLA evaluation suites

See it in our Sample Lab

SignIQ capability

We replay your policy in our cell, capture successes and labelled failures with operator-rated severity, and ship them in the same schema as our teleop bundles. Per-frame reward and outcome labels included.

Simulation & Synthetic Trajectories

Available now

Demonstrations generated in simulators with domain randomization, ground-truth segmentation, and infinite scale — paired against real captures for sim-to-real transfer.

Sample
1 live · interactive
Export
LeRobot v2 · GR00T-LeRobot · …

Signals

RGBDepthSegmentationActionStateObject pose

Used for

PretrainingPost-trainingFine-tuningEvaluation

Models that train on this

GR00T · OpenVLA sim eval · Sim-to-real diffusion policies

See it in our Sample Lab

SignIQ capability

Hybrid sim/real collection: digital twins of our physical scenarios with matched control frequencies. Sim and real episodes share an embodiment tag and modality map for clean co-training.

Human Egocentric & Activity Video

Available now

First-person and third-person human task video used for affordance priors, semantic skill learning, and world-model pretraining without robot actions.

Sample
3 live · interactive
Export
LeRobot · MP4 + JSON · …

Signals

RGBAudio (optional)Hand poseObject interaction

Used for

PretrainingWorld Model

Models that train on this

VLA reasoning pretraining · World models · Affordance & skill priors

See it in our Sample Lab

SignIQ capability

Operator-recorded egocentric capture with 150-D hand pose, finger angles, IMU, gesture probabilities, and dense subtask spans. Consent-cleared, face-blurred, PII-reviewed before delivery.

Human Motion Capture & Retargeting

Available now

Skeleton/joint motion capture retargeted to robot morphology — the foundation for whole-body humanoid locomotion, balance, and manipulation posture priors.

Sample
2 live · interactive
Export
LeRobot · GR00T-LeRobot · …

Signals

Joint anglesRoot motionBody keypointsRetargeted action

Used for

PretrainingLow-level Control

Models that train on this

Whole-body humanoid controllers · Locomotion / balance priors · Hierarchical VLAs

See it in our Sample Lab

SignIQ capability

Markered and markerless mocap with embodiment-specific retargeting pipelines for ALOHA, GR1-class humanoids, and dexterous hands.

Force / Torque & Tactile Signals

Available now

Six-axis force/torque, fingertip tactile arrays, and contact event traces. Essential for contact-rich manipulation, insertion, and dexterous skills.

Sample
1 live · interactive
Export
LeRobot · HDF5 · …

Signals

6-axis F/TTactile arrayContact onsetSlip events

Used for

Fine-tuningLow-level Control

Models that train on this

Contact-rich diffusion policies · Dexterous BC · Whole-body humanoid manipulation

See it in our Sample Lab

SignIQ capability

Wrist F/T at 100 Hz, fingertip tactile arrays, and contact-event tagging synchronized to RGB and proprioception.

Language & Semantic Annotations

Available now

Episode-level instructions, subtask spans, hindsight captions, and failure-mode descriptions — the supervision signal for VLA instruction following.

Sample
1 live · interactive
Export
LeRobot · JSON · …

Signals

Task instructionSubtask spansHindsight captionsSemantic action labels

Used for

PretrainingFine-tuning

Models that train on this

RT-2-style VLAs · OpenVLA instruction following · Hindsight VLM relabeling

See it in our Sample Lab

SignIQ capability

Three-tier labels: operator-written instruction, hindsight VLM captions, and reviewed subtask spans aligned to action chunks.

QA & Reliability Metadata

Available now

Sync drift, missing-frame counts, joint-limit violations, calibration provenance, and shape consistency — the hidden layer that makes training reliable.

Sample
11 live · interactive
Export
JSON (per-bundle health.json)

Signals

Sync driftFrame integrityJoint limitsCalibration filesSchema validity

Used for

PretrainingFine-tuningEvaluation

Models that train on this

All — used as training-time filters and per-episode weighting

See it in our Sample Lab

SignIQ capability

12-check automated QA pipeline runs on every episode before delivery; failures are re-collected, not patched.

See the data, not just the spec sheet.

The Sample Lab opens a real episode end-to-end — synchronized cameras, action vs. state, end-effector trajectory, force/torque, language, and dataset health — backed by an actual SignIQ teleoperation capture.