Real-world training data for physical AI
We collect, narrate, annotate and deliver the multimodal datasets that teach robots to act. Captured across real industrial, clinical, domestic and field environments, and quality-checked before it ever reaches your model.
Ego array · industrial assembly · multi-view
Capturing
At a glance
A quality-checked library you can license today, plus live capture crews that stand up a bespoke program and scale it to a thousand hours a day.
Hours of OTS data
Off-the-shelf and quality-checked. Ready to license now.
Hours narrated
Industrial work, captured with synchronized audio and transcripts.
Hours per day
Live collection capacity across our active regions.
Countries
India, Latin America, Europe and Africa.
License from a cleared, documented library and start training immediately.
Commission capture for your exact tasks, environments and edge cases.
Mono and stereo video, narrated audio, depth, IMU and dense labels.
Consent-based sourcing aligned to the law of every region we work in.
Why now
Humanoids are walking onto factory floors. Autonomous systems are moving through warehouses, roads and hospitals. These systems do not learn from clean theory; they learn from messy reality. The bottleneck has shifted from algorithms and hardware to the real-world data that teaches a robot to act.
Robotics programs are crossing from demos into deployments where reliability is non-negotiable.
Vision-language-action and world models need diverse, action-grounded demonstration data.
Teams can capture far fewer real hours than their models actually need to generalize.
The problem
Why robotics data is hard
What it costs teams
Without edge-case and failure coverage, a policy that looks great in the lab breaks in the field.
Time that should go to the model goes to capture, cleaning and labeling instead.
Misaligned data quietly degrades a policy in ways that are painful to debug later.
What we do
One partner for the whole job: define the task, capture it in the real world, narrate and annotate every action, check the quality, and deliver it straight into your pipeline.
End-to-end partner
Signal types captured
Real environments
Stage QA pipeline
The pipeline
Step 1
Multimodal capture across real environments.
Step 2
Operators describe each action, on the clock.
Step 3
Dense, action-grounded labels and language.
Step 4
Five-stage QA on every batch.
Step 5
Pushed straight into your training stack.
Input
Raw, real-world activity in.
Output
Training-ready datasets out.
Traceable
Every stage is versioned and auditable, so data is traceable from capture to delivery.
Capabilities
Capabilities · 1 of 4
Purpose-built capture programs, designed around the behavior you want your model to learn. We agree on the tasks, environments and edge cases up front, then capture them at scale across synchronized cameras and sensors.
Frame-accurate alignment across every sensor.
Built around your capability and benchmark goals.
Environments and operators that generalize.
Timestamps, activity labels and descriptors.
Capabilities · 2 of 4
Dense, robotics-grade labeling that goes well beyond boxes on images. We annotate the sensor streams, actions and language that vision-language-action models actually consume, and a human expert reviews every batch.
Aligned in time to what the operator did.
Vision, language and action linked together.
Stable label schemas across every batch.
Expert reviewers keep accuracy high.
Capabilities · 3 of 4
The gold standard for demonstration data. Skilled operators drive robots through target tasks and produce the exact state-action pairs that imitation learning consumes, with no inverse dynamics and no reward engineering.
Directly consumable by imitation learning.
Scaled without sacrificing signal.
Mapped to your evaluation up front.
Ongoing drops to evaluate policies.
POST /v1/datasets/export 200 OK
{
"dataset": "industrial_assembly_v3",
"format": "lerobot | rlds | coco",
"signals": ["rgb","stereo","audio","imu"],
"delivery": "stream",
"qc_score": 96.4
} Capabilities · 4 of 4
Data should not arrive as a zip file over email. Our delivery APIs push training-ready data straight into your environment, with the structure and metadata your tooling already expects.
10+
Export formats
3
Cloud targets
<1s
Stream latency
What we capture
Every clip we deliver is more than video. It is a stack of synchronized signals, aligned to the frame, so perception and vision-language-action models can consume it directly.
Single-camera egocentric view at 1080p, 30 fps.
1080p · 30 fps · egocentric
Synced stereo pairs with metric disparity and depth.
65 mm baseline · metric depth
Operators describe each action, with clean transcripts.
Speech + timed transcript
6-axis IMU at 300 Hz, synced to the camera to 1 µs.
6-axis · 300 Hz · 1 ms sync
Hand keypoints, grasp points and contact events.
Keypoints · grasp · contact
Tasks split at the moment the operator’s intent changes.
Intent-boundary segments
Objects detected, labelled and located in frame.
Detect · label · locate
Scene geometry, multi-view and environment tags.
Multi-view · scene tags
Capture hardware
Lightweight, cap-worn capture units that an operator can wear through a full shift. Choose a single-camera mono rig for reach and volume, or a calibrated stereo rig when your model needs metric depth.
Specification
Stereo rig
Mono rig
Camera configuration
2× RGB stereo
1× RGB
Per-camera resolution
1920×1080 @ 30 fps
1920×1080 @ 30 fps
Native FOV
120° H / 80° V / 150° D
120° H / 80° V / 150° D
Post-rectification FOV
118.17° / 61.47° / 143.76°
—
Baseline
65 mm
n/a · single camera
Shutter
Global
Global
IMU
6-axis @ 300 Hz
6-axis @ 300 Hz
Camera-IMU sync
1 µs
1 µs
Recording
On-device / SD card
On-device / SD card
Power
USB power bank
USB power bank
Weight
127 g
~95 g
Form factor
Integrated cap-worn
Integrated cap-worn
Stereo values are production spec. Mono rig figures marked “—” or “~” are indicative; final mono specifications available on request.
1 µs
Cam-IMU sync
300 Hz
IMU rate
127 g
Stereo weight
Industrial expertise
Operators narrate each task aloud as they perform it, so every clip ships with synchronized audio and a clean transcript. Your model learns the intent behind the motion, not just the motion.
10,000+
Narrated hours
Audio
+ transcript per clip
60 fps
High-speed lines
Factory · pipe-fitting · narrated
Experience
Industrial · 60 fps
Skilled trades · lathe
Clinical · 3D CT
Clinical · in-vivo
Embodied · Unitree G1
Domestic · hand pose
Field · outdoor
Construction · heavy
Sourcing
Five ways to get the data you need, from a library that is ready today to a crew that captures the exact scenario your model keeps failing on. Every route is consent-based and documented.
Trained operators capture on location in each region, scaling to 1,000+ hours a day.
Every participant opts in. Consent, rights and usage are recorded per clip.
Factories, workshops and clinics captured under formal agreements.
Repeatable rigs for targeted tasks, edge cases and benchmarks.
250,000+ hours already collected, cleared and ready to license.
We go and capture the rare scenarios that your current data misses.
Speed
Because our crews, rigs and library are already in place, a new program moves quickly. You see a validated pilot batch in about two weeks, then we ramp capacity from there.
01
We align on tasks, environments, edge cases and data rights.
Day 0
02
Devices configured and operators briefed in the target region.
Week 1
03
A validated pilot batch lands in your pipeline for review.
~Week 2
04
Capacity ramps toward a thousand hours a day.
Ongoing
05
Fresh drops as your model and benchmarks evolve.
Continuous
Quality
Bad data fails silently, so we build verification into every stage. Issues surface before data reaches your model, and every batch stays auditable from source to delivery.
Acceptance criteria agreed up front and checked on every batch.
Alignment verified across modalities and action boundaries.
Skilled reviewers verify labels a model alone would miss.
Full provenance from capture to delivery for audit and rollback.
Trust & compliance
Faces, plates and other identifiers are automatically redacted before data leaves our pipeline.
Every contributor opts in; rights recorded per clip.
Automatic face and identifier blurring on capture.
Secure transfer and storage, access-controlled.
Regional handling aligned to local law.
Every dataset versioned and auditable.
Contributors can revoke; data is traced and removed.
Global reach
A distributed footprint gives your models the environmental and cultural diversity they need to generalize, each region captured under its own data-protection regime.
DPDP Act 2023
LGPD · regional law
GDPR · UK GDPR
POPIA · NDPR
Partners
We work with the teams building physical AI, from early-stage startups to the labs setting the frontier. Client names are withheld under NDA; references are available on request.
Early teams that need production-grade data without building a collection org first.
Groups training large multimodal and world models on real-world signals.
Manufacturers and integrators shipping manipulation and mobile platforms.
Groups that need teleoperation and action-grounded demonstration data.
Clinical capture handled with the privacy and rigor the domain requires.
Perception and navigation teams across field, logistics and AV programs.
Differentiation
01
Industrial, clinical, domestic and field footage from the places robots actually work.
02
Audio and transcripts carry the intent behind every motion, aligned to the frame.
03
Mono and stereo rigs, sensor-synced to the microsecond, built for long field shifts.
04
License 250,000+ ready hours today, or commission exactly what you are missing.
05
Defined QA gates, expert review and full provenance on every batch.
06
Sourced and handled to the law of every region we operate in.
FAQ
dexset delivers ML-ready robotics training datasets: real-world video captured to your task brief, annotated with objects, actions, poses, task states, and failures, validated for coverage and consistency, and exported in COCO, YOLO, Pascal VOC, JSON, CSV, or a custom schema mapped to your pipeline.
Generic vendors label footage you already have. dexset runs the full loop: task definition, real-world capture with trained operators, robotics-specific annotation, quality scoring, and delivery. Labels carry physical context — grasp points, task states, failure markers — that general-purpose labeling teams routinely miss.
No — we complement it. Simulation is excellent for scale and rare-state coverage, but deployment requires real-world variation: lighting, clutter, occlusion, and human behavior. Most teams use dexset data to fine-tune, validate sim-to-real transfer, and build held-out evaluation sets.
A pilot of under ten hours of captured data typically ships in two to three weeks, including annotation and quality review. Larger programs run as recurring batches, so your first training-ready data arrives early rather than at the end of the engagement.
You do. Custom-captured datasets are delivered under exclusive ownership or exclusive license terms agreed before capture begins. Every clip carries source, consent, and capture records, so provenance is auditable from raw footage through to the exported training file.
Let's work together
Tell us what you’re training and we’ll bring a data plan. A dexset data expert replies within one business day.
Web