Skip to content
Solutions · Physical AI & Robot Data

Physical AI Training Data: Multimodal Field Capture for Robot Learning

Robot-ready physical AI training data with the full multimodal stack — calibrated stereo depth, 16x16 tactile force, 200Hz IMU and 21-point hand pose, synced on one clock rather than mono RGB. Real human demonstrations, consented at the source and delivered LeRobot-ready.

Book a Physical AI Data Strategy Session Free 30-minute call · mutual NDA included
12Synchronized modalities per session — where most datasets ship one.
8Ground-truth layers on every episode, synced on one clock.
99%Human-verified before delivery, so you can train the moment it lands.
Outcomes

What's in Our Physical AI Dataset

Every episode is one synchronized multimodal bundle — far beyond the mono-RGB video most corpora ship. Eight ground-truth layers on every episode.

Stereo Depth, Not Just RGB

Metric per-frame depth with intrinsics and extrinsics — real trajectories, not scale-ambiguous projections.

Dual Wrist Cameras

Left and right wrist views — the close-up a robot actually sees at execution time.

Tactile Force Arrays

16x16 taxels at 100Hz per finger — the contact signal simulation can't fake.

IMU at 200Hz

Head and both wrists — fast motion that 30Hz video aliases away.

21-Point Hand Pose

Per-frame keypoints, retargetable to grippers and dexterous hands.

Action Segments + Tracks

Human-verified verb-noun segments and persistent object tracks.

The Problem

Why Capture-First, Not Scraped or Synthetic

Scraped web video is mono and scale-ambiguous. Synthetic can't fake contact, friction or slip. We capture first-hand on real floors, so depth, force and motion are measured, not inferred — and every episode carries consent and provenance you can audit. Three gaps every scraped or synthetic corpus leaves open:

1 Scraped web video is mono and scale-ambiguous — no metric depth, no reliable trajectories to learn from.
2 Synthetic data can't fake contact, friction or slip — the tactile signal a manipulation policy actually needs.
3 No consent or provenance — a licensing and audit liability the moment you train a commercial model.
The Capture-First Answer

We measure the physical signal, not infer it.

Metric depth + tactile + 200Hz IMU on one synced clock — the multimodal signal a policy needs, and the one mono RGB can't provide. Real human demonstrations, captured on real floors, consented at the source.

Metric depth, not scale-ambiguous projections
Real tactile force, not synthetic guesses
Consent and provenance on every episode
Multimodal Beyond RGB

The 8 Ground-Truth Layers on Every Episode

Most capture is mono RGB from gig workers or lab teleop. Ours pairs stereo depth, dual wrist views, tactile and 200Hz IMU on one clock, across factories, workshops and specialty settings others can't reach.

1

Metric stereo depth on every frame

Calibrated per-frame depth with intrinsics and extrinsics — real, scale-accurate trajectories a policy can learn from, not the scale-ambiguous projections you get from scraped mono video.

2

Dual wrist camera views

Left and right wrist views give the close-up perspective a robot actually sees at execution time — the manipulation context that a single head-mounted RGB feed can't capture.

3

Tactile force at 16x16 taxels

16x16 taxel arrays at 100Hz per finger record grasp force, slip and contact — the signal simulation can't fake and scraped video never sees.

4

200Hz IMU on head and both wrists

High-rate inertial capture on the head and both wrists preserves the fast motion that 30Hz video aliases away — critical for contact-rich and dynamic tasks.

5

21-point hand pose keypoints

Per-frame 21-point hand keypoints, retargetable to parallel-jaw grippers and dexterous hands — so the same episode maps into your robot action space.

6

Action segments and object tracks

Human-verified verb-noun action segments and persistent object tracks label what happened and to what — grounded, language-conditioned supervision for VLA fine-tuning.

7

Consent and SHA256 provenance

Every episode carries a signed release, site agreement and SHA256 chain of custody — commercial-training-safe data you can audit, not a scraped corpus of unknown origin.

8

One synced clock across all layers

Depth, tactile, IMU, hand pose, action segments and tracks are all synchronized on a single clock, so cross-modal alignment is exact — no post-hoc guessing at timestamps.

Start Today

Evaluating Data Sources for Your Policy Stack?

Send us the manipulation task. We'll map the capture and send a representative sample pack — the same modalities, calibration and delivery format you'd get in production, so your team can inspect sensor alignment before any commitment.

Book a Strategy Session →
Ask us about
Multimodal capture — stereo depth, tactile, 200Hz IMU and hand pose on one clock
Fine-tuning OpenVLA, π0 or GR00T on real human demonstrations
LeRobot, RLDS/TFDS and ROS 2 / Foxglove MCAP delivery formats
How our human-demonstration data complements Open X-Embodiment
Custom capture to your task taxonomy, sites, objects and acceptance criteria
Commercial licensing, consent releases and SHA256 provenance
Own the Capture

When You Need Custom Capture to Your Taxonomy, Not a Generic Corpus

Off-the-shelf robot corpora cover common pick-and-place well. But teams pushing real physical AI need capture that a scraped or synthetic dataset structurally can't deliver:

Your exact tasks, environments and objects — captured to spec, not the median household kitchen.
Metric depth and tactile force measured first-hand — not inferred from scale-ambiguous scraped video.
Capture on real floors others can't reach — factories, workshops and specialty settings.
Ground truth on one synced clock — action segments, hand pose, tracks, depth, tactile and IMU.
Consent and provenance you can audit — signed releases and SHA256 chain of custody per episode.
Exclusive licensing on custom episodes — the data becomes a moat, not a shared corpus.

Send your task taxonomy and we scope a capture program against it — sites, objects, action set, episode counts and acceptance criteria — so the data you train on is your data, measured on real floors, not a generic corpus everyone else already has.

Questions

Frequently Asked Questions

Every episode is captured under a signed release and site agreement that permit commercial model training, with a per-episode chain of custody you can audit. Licensing is per-dataset and non-exclusive by default; exclusive and custom terms are available. We'll send the actual licence text before you commit.

Yes. Datasets ship LeRobot-ready and load with LeRobotDataset("neuralchainai/<dataset>"), and are available as RLDS/TFDS or ROS 2 / Foxglove MCAP on request — no conversion project before you can train.

Open X-Embodiment is robot trajectories; ours is human demonstration data with far richer sensing (metric depth, tactile, 200Hz IMU) and broader diversity. They're complements — most teams train on both and budget for the retargeting step between our hand pose and a robot action space.

Yes — it's the main reason teams come to us. Send your task taxonomy and we scope a capture program against it: sites, objects, action set, episode counts and acceptance criteria. Exclusive licensing on custom episodes is available.

Start with the sample pack — a small but fully representative delivery with all modalities and calibration, in LeRobot and MCAP form, so your team can inspect sensor alignment before any commitment. Request it on a strategy call.

Eight synced ground-truth layers — action segments, 21-point hand pose, object tracks, depth, tactile, IMU, consent and SHA256 provenance — arrive validated on one clock. Load a sample and evaluate it in an afternoon.

See the Data Before You Decide

A 30-minute strategy call. We'll walk through your manipulation tasks, target model and delivery format — then scope a representative sample pack so your team can inspect sensor alignment and evaluate it in an afternoon.

Discuss Your Project