Skip to content
Solutions · Physical AI & Robot Data

Embodied AI Training Data: Demonstrations for Robot Policies & VLAs

Embodied AI training data for robot policies and vision-language-action models: real human demonstrations with synced depth, tactile, dual-wrist views and 21-point hand pose, action-labelled and LeRobot-ready. This is embodied AI in the robotics sense, not embodied cognition.

Book an Embodied AI Data Strategy Session Free 30-minute call · mutual NDA included
21Hand-pose keypoints tracked per hand, retargetable to grippers and dexterous hands.
2xWrist cameras, left and right — the close-up a robot has at execution time.
8Ground-truth layers on every episode, synced on one clock.
Outcomes

What's in Our Embodied AI Dataset

Every episode is one synchronized bundle for training robot policies and VLAs: multi-view video, metric depth, tactile, IMU and per-frame action labels.

Action-Labelled Demos

Per-frame action labels retargetable to a robot end-effector.

Depth and Tactile

Metric stereo depth and 16x16 tactile force, beyond mono RGB.

Dual Wrist Views

Left and right wrist cameras, the view a robot has at execution.

Hand to Gripper Mapping

21-point hand pose that retargets to grippers and dexterous hands.

Task Diversity

Real factory, workshop and service tasks, not tabletop toys.

Failure and Correction

Real mistakes and recoveries, the data teleop sets rarely contain.

The Problem

Why Human Demonstrations, and Where They Fit

Human demonstrations scale far cheaper than teleop and reach environments a robot can't, but they carry an embodiment gap: a human hand is not a parallel gripper. We make that gap tractable with metric depth and 21-point hand pose, so reach, grasp and place retarget cleanly. In-hand reorientation is the honest hard case.

1 Human demonstrations carry an embodiment gap — a human hand is not a parallel gripper, so raw video alone doesn't transfer to a robot.
2 Mono web video is scale-ambiguous, and lab teleop cannot reach the factories, workshops and service settings real tasks live in.
3 Sim benchmarks under-measure contact, and teleop sets rarely contain the failures and corrections a policy has to learn from.
Where They Fit

We make the embodiment gap tractable.

Depth, tactile, dual wrist and 200Hz IMU on one clock: the multimodal signal a policy needs, and the one mono video can't provide. Metric depth and 21-point hand pose retarget human reach, grasp and place cleanly to an end-effector.

Metric depth and 21-point hand pose for clean retargeting
Tactile and 200Hz IMU on one synced clock
Real failures and corrections, consented at the source
Where It Fits

Where Our Data Gets Used

Every episode ships LeRobot-ready, with RLDS/TFDS and ROS 2 / Foxglove MCAP on request, so it loads straight into an OpenVLA, pi0 or GR00T fine-tune.

1

Fine-tuning a VLA

OpenVLA, pi0 or GR00T on real, language-conditioned demonstrations.

2

Imitation learning

Human video with action labels for behaviour cloning and policy learning.

3

Contact-rich manipulation

Grasp, slip and insertion, backed by the tactile and depth modalities mono video never sees.

4

Custom task taxonomy

Your eval's exact tasks and environments, captured to spec at partner sites rather than a generic corpus.

5

Managed capture where the tasks live

We run consent-first capture at partner sites — factories, workshops, repair bays and craft studios that gig marketplaces and lab teleop cannot reach — with provenance you can audit.

6

Delivery into your training stack

Datasets ship LeRobot-ready and load with LeRobotDataset, with RLDS/TFDS and ROS 2 / Foxglove MCAP on request, so there is no conversion project before you can train.

7

Retargeting human motion to your embodiment

Metric depth and 21-point hand pose retarget human reach, grasp and place cleanly to a parallel gripper or dexterous hand — the retargeting step raw video alone can't support.

8

Synced multimodal ground truth

Depth, tactile, dual wrist and 200Hz IMU on one clock — the multimodal signal a policy needs, and the one mono video cannot provide.

Start Today

Fine-tuning a Policy or VLA?

Tell us the tasks and embodiment. 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
Fine-tuning OpenVLA, pi0 or GR00T on real human demonstrations
Retargeting human hand pose to a parallel gripper or dexterous hand
Metric depth, 16x16 tactile and 200Hz IMU synced on one clock
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 and acceptance criteria
Own the Capture

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

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

Your eval's exact tasks and environments — captured to spec at partner sites, not tabletop toys.
Metric depth and 21-point hand pose — so human reach, grasp and place retarget cleanly to an end-effector.
Capture on real floors others can't reach — factories, workshops, repair bays and craft studios.
Failure and correction episodes — the real mistakes and recoveries teleop sets rarely contain.
Consent and provenance you can audit — signed releases and a per-episode chain of custody.
Action-labelled, language-conditioned demos — grounded supervision for behaviour cloning and VLA fine-tuning.

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, captured where the tasks actually live, not a generic corpus everyone else already has.

Questions

Frequently Asked Questions

Every episode is captured under a signed release permitting commercial training, with an auditable per-episode chain of custody. Non-exclusive by default; exclusive and custom terms available.

Yes. Datasets ship LeRobot-ready and load with LeRobotDataset, and are available as RLDS/TFDS for OpenVLA-style pipelines or ROS 2 / MCAP on request.

Through retargeting. Metric depth and 21-point hand pose map human motion to an end-effector for reach, grasp and place. Budget for the retargeting step, which is why we ship the depth and pose that make it work.

Start with the sample pack, a representative delivery with all modalities and calibration, so your team can inspect sensor alignment first.

No. Embodied AI here means AI that perceives and acts in the physical world, robot policies and VLAs. Embodied cognition is a separate psychology and philosophy theory.

This is human demonstration data, so it complements a robot-trajectory corpus like Open X-Embodiment rather than replacing it. Most teams train on both and budget for the retargeting step between our hand pose and a robot action space; note that sim benchmarks under-measure the contact modalities.

See the Data Before You Decide

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

Discuss Your Project