LiDAR SLAM
DeployedA point cloud of the street builds around the operator as they walk, with their exact position and heading tracked through it, even between tall buildings.
We train robots in the real world
Precision, high‑fidelity multimodal ground truth data corpus for Robotics RL and Real World Deployment
Telemety records people doing real work on New York streets and delivers it as robot-ready training episodes: full-body human multimodal capture, time-aligned, faces blurred before the data leaves our pipeline, published as ROS 2 bags.
Home of Physical AI + for the Real World and Beyond
3 Laws SafeFull body
human multimodal capture system
Collecting
on New York streets, in every season
ROS 2
topics and bags, rosbag2 / MCAP
All-weather conditions
Multi-Rig Outdoor 24 Hour Capture
We study locomotion and obstacle avoidance during rain, snow and sunshine, year round.
Capture log
Already recording in New York City
New York is arguably the most complex urban environment in the United States. Broken curbs, heavy doors, scaffolding, crowds that do not yield. If a robot can learn to navigate this city, it has a chance anywhere.
Latest session · 1 Nov 2026 · Midtown · full-body multimodal
08:00 commute
12:00 parks and stairs
16:30 evening rush
Contributor programme
Telemety EgoCentric Capture app
Our egocentric capture app is supplied to approved contributors only. If your work already takes you through the city on foot, you can record it with us and be paid for the hours we approve.
h
Total hours of capture
PB
Total data captured
Currently open to city walkers · urban cleaners · construction workers · emergency response professionals
Live estimate, accruing from the capture schedule since 3 Nov 2026.
THX H2R rig
Not for sale: available for leased capture
The THX H2R capture system is not sold as hardware. It is deployed on leased capture engagements: our rig, our trained operators, your routes and your tasks, with the resulting episodes delivered to you.
Sessions logged
PB
Total data captured
How it works leased by the rig-day · ten-day minimum · same pipeline and formats
Live estimate, accruing from the capture schedule since 3 Nov 2026.
See capture termsContributor application
Tell us what you do and where you work. Applications go to team@telemety.ai, we reply to every one.
Your application, send it to team@telemety.ai
Nothing has been sent yet. Send the email, or paste this to team@telemety.ai.
Why real streets
Robots trained in labs and simulation struggle in the real world
Most robot training data is video and pose. We record people doing real work on real streets, capturing what they see and feel, where they look and exactly where they are, then turn it into training data robot teams license.
Labs and simulation
- Video and pose
- Predictable scenarios
- Clean floors, known lighting, no weather
- Pedestrians modelled, not observed
On real streets
- Full-body multimodal capture, time-aligned
- Broken curbs, heavy doors, scaffolding
- Snow, rain, night, crowds
- Every sensor on one clock
Modalities
Recorded, not rendered
Coloured SLAM
DeployedEvery point in the cloud carries the colour of the surface it came from. The result is a 3D map that looks like the street it came from.
Gaze and intent
DeployedWhere the operator was looking, on every frame, so each recording shows the attention that preceded an action and where they were about to go.
3D Hand Pose Kinematics
DeployedA full skeleton for each hand on every frame, in three dimensions and with a confidence score per hand. It is what turns “the person picked it up” into a grasp a manipulator can be trained on: finger pose, wrist orientation, approach and release, recorded on real work rather than demonstrated in a lab.
360° video
DeployedThe whole scene around the operator is recorded at once, so pedestrians, traffic and storefronts on every side are captured, not only what is straight ahead.
Full-body motion capture
DeployedThe operator's full skeleton as they walk, on the same clock as every other stream, for a whole session.
Egocentric video
DeployedExactly what the operator sees, including their own hands as they reach for, grip and handle objects in the world.
THX H2R capture system
Full Body Human Multimodal Capture System
CollectingNew York City · all weather · year round
Recording now, on real streets, with the people who already do the work. The system is not for sale; it is deployed on leased capture engagements, and its specification is shared with customers on request.
Product · Included with the THX H2R capture system
OmniFusion Data Dashboard
for Real World Robotics Training and Deployment
The software half of THX H2R. Every stream the rig records arrives here, time-aligned and playable: configure what a session captures, watch it live, annotate it, debug it, and export it in the format your stack already reads. Our own models run in real time as filters on the incoming data, each one fine-tuned on our own recordings.
01
Capture module configuration
Turn capture modules on and off per session, so a capture is scoped to the task you are buying it for.
02
Playback
Scrub a whole session with every stream frame-aligned on the same millisecond clock, live or after the fact.
03
Annotation
Label events, objects and intent in place, against the taxonomy that ships with the episode.
04
Debugging
Sync drift, dropped frames and calibration faults are surfaced per stream, so a bad session is caught before it is delivered.
05
Real-world map and location
The route on the map as it was walked, anchored in the world.
06
Surround vision
The full 360° field around the operator, not only the forward view, with the model filters drawn over it.
07
Session logs
Every session recorded, searchable by task, route, surface, weather or time of day.
08
Custom AI agent integration
Point your own agent at the corpus and let it query, filter and pull episodes through the same interface.
09
Multi-format export
ROS 2 topics and bags, rosbag2 and MCAP, with other formats on request.
Privacy by design
Anonymised before the data leaves our pipeline
Street capture meets a buyer's legal team exactly once. We built for that review from the first session rather than after it.
01
Faces blurred at source
A blurring model runs inside the pipeline, before an episode is written to the archive or delivered. No biometric templates are produced or stored.
02
Conversations muted
Speech is removed. What remains is what a robot has to hear: sirens, horns, traffic, and short pedestrian commands such as "excuse me".
03
Consent where people take part
Anyone who interacts with an operator as part of a recorded task signs a consent and release that is assignable to licensees.
04
Encrypted end to end
Episodes are encrypted in the archive and on delivery. Licence terms, retention periods and the pipeline description are available for your counsel.
Processing and delivery
Episodes you can load the day they arrive
Every session goes from raw sensor streams to clean, labelled data, processed by models we train ourselves on our own recordings, and published in the format your stack already reads.
01 / CLEAN
Published in ROS 2
ROS 2 topics and bags, rosbag2 and MCAP, with sync verification, denoising and calibration on every stream.
02 / MODELS
Our own models
Models trained in-house on our own recordings run across every stream, on the edge and in post.
03 / GROUND
Traversability
Our algorithm maps the ground a robot can safely cross, scored per segment and attached to the trajectory.
04 / LABELS
Annotated
Boxes, segments, events and intent labels, with a taxonomy you can slice by task, surface, weather or time of day.
Try one episode
An anonymised sidewalk episode with every deployed stream, as an MCAP bag. All formats are ready for integration into your own or third-party partner platforms. Tell us the task you care about and we will match the slice.
Request the sample episode# deployed streams, faces blurred ros2 bag play telemety_midtown_2026-10-28.mcap # or hand the same bag to your own pipeline mcap info telemety_midtown_2026-10-28.mcap
Product · Telemety H2R Platform
Human‑2‑Robot Platform
for Human-2-Robot Data Collection, Training and Real World Deployment
Where the data becomes a policy. View your sessions, build and tune robot policies against them, and train for the machine you are actually deploying: your own custom build or an existing platform of any form factor. The data and the platform are agnostic by design: no proprietary container, no tie to one vendor, one simulator or one morphology.
Proof of concept · simulation study
Human ground truth makes a humanoid walk better
What the loop is for. We took the centre-of-pressure path a person actually walks, measured with Telemety pressure insoles, retargeted it onto a humanoid’s gait in simulation, and ran it against engineered baselines over paired courses — the same human-to-robot path the platform is built around, tested end to end.
+16.4%
wider foot-support margin at the foot edge than an engineered linear CoP path — 12 paired courses, Holm p = 0.008
60%
the human path compressed to 60% of its excursion and still inside the non-inferiority margin on falls — 128 paired fast courses
64
taxels per sole, resolving foot contact an ankle force/torque sensor never sees
42–52%
lower support-fraction error than a 6-axis F/T sensor on unknown rough terrain
Target mission domains
Fire rescue and firefighting
Heat and smoke detection, hose and tool handling, victim search, toxic gas handling, structural judgement.
Search and rescue in rubble
Rubble mapping, victim detection, debris removal, safe extraction, medical stabilisation.
Traffic and crowd control
Traffic flow reading, signalling by hand, accident response, crowd guidance, emergency routing.
Emergency response
Situational awareness, civilian protection, evacuation support and the judgement calls that go with them.
These are the domains the capture programme and the platform are aimed at. None is a shipping capability today; each begins as recorded human experience of people who already do the work.
01 / VIEW
Your data
Every session you have commissioned or licensed, every stream, time-aligned and playable in the browser.
02 / BUILD
Policies
Build and tune robot policies directly against your own captures, rather than against someone else's demonstration set.
03 / RETARGET
Robot library
Train for your own custom robotic builds or existing robots across form factors, humanoids, quadrupeds, rovers and drones.
04 / DEPLOY
Real world
Take the policy to the street it was recorded on, and bring the next session back in through the same loop.
Format agnostic
ROS 2 topics and bags, rosbag2 and MCAP, with other formats on request. Nothing is locked in a container only we can open.
Robot agnostic
Human motion is retargeted to the kinematics of the machine you name, not to a single reference platform.
Stack agnostic
Use our platform end to end, or take the episodes and run them in your own or a third-party partner platform.
Access to the H2R Platform is granted per account and is included with a pilot. Tell us your target platform and we will confirm retargeting support before you commit.
Dataset taxonomy
Telemety Intelligence Library
Every session we record is indexed against one taxonomy, so a corpus becomes something you can slice rather than something you have to sift. The Library is that taxonomy published three ways: by the sector you are deploying into, by the capability you are training, and by the world you are reconstructing.
A
Curated datasets by sector
Task-matched slices for the environments robots are being sent into first.
Last Mile Logistics
Doorways, lobbies, stairwells, handovers and the full first-to-last-foot of a delivery, from a courier's own point of view.
RecordingDisaster Response & Urban Rescue
Obstructed routes, debris, confined access and degraded footing, with the decisions a responder makes in them.
In planningRetail & Customer-First Robotics
Shelf and aisle navigation, queueing, approach distance and the manners of serving a person who did not ask to be served.
In planningHospitality & Service
Trays, doors, lifts, tight interiors and the handover etiquette that decides whether a service robot is tolerated.
In planningHigh-Rise Construction
Site footing, material handling, hoists and scaffold, captured where the surface changes hour to hour.
In planningUrban Maintenance
Kerbs, covers, standpipes, cellar doors and the long tail of street furniture a maintenance robot has to work around.
In planningPublic Safety & Traffic Control
Crossings, signals, crowd flow and vehicle interaction at the junctions where pedestrian and traffic behaviour collide.
In planning
B
For robotics companies: the data matrix
The same corpus indexed by the capability you are training, and the streams that supply the ground truth for it.
| Capability | What it covers | Streams that feed it |
|---|---|---|
| Locomotion | Gait, balance, stairs, kerbs, slopes and recovery on uneven ground | mocap · IMU · insoles · LiDAR odometry |
| Dexterity | Reach, grasp, handover, doors, handles and two-handed tasks | egocentric · hand pose · tactile gloves |
| Perception | What was visible, from where, at what range and in what light | 360° RGB · egocentric · stereo depth · LiDAR |
| Navigation | Route choice, crowd negotiation, yielding and traversable ground | fused LiDAR / GNSS trajectory · traversability |
| Intent | Where attention went in the seconds before an action | gaze · head pose · egocentric |
| Interaction | Handovers, queueing, approach distance and short spoken commands | 360° RGB · egocentric · consented audio |
| Force & contact | Ground reaction, load transfer and contact pressure | insoles · tactile pads · IMU |
Streams drawn from families marked Integrating are not yet in every episode. The stream status table above is authoritative.
C
For world model companies
The reconstruction layer: the city as geometry, appearance and labelled objects rather than as episodes.
01
Digital assets
Point clouds and coloured reconstructions of real blocks, built from the same sessions rather than modelled after the fact.
02
Mapping
Routes anchored in the world, repeatable across sessions and seasons.
03
Rich object classes
The long tail a synthetic city never includes: scaffolding, cellar doors, standpipes, bollards, plate steel, delivery carts, street dining.
04
Conditions
The same geometry under rain, snow, night and glare, so appearance and structure can be separated.
Sectors marked In planning are specified and not yet recorded. Tell us which one you need and we will scope a collection against it rather than sell you an adjacent slice.
In development · 2027
Score a route before you send a robot down it
The Urban Robotics Deployment Score turns our recordings into a judgement about a specific machine on a specific street: clearance, surface, crowding, crossing risk, and the cost of getting it wrong.
Per segment
Traversability and risk scored block by block from recorded sessions
Digital twin
Routes rehearsed against a twin of the blocks we have recorded
For
Sidewalk operators choosing routes, cities writing permits, insurers pricing them
Build with us
Tell us what your robot has to do.
Send the task, the platform and the environment. We will tell you whether a slice of the corpus already covers it, what a pilot would contain and when we could deliver it. If we cannot help, we will say so.
We answer technical questions technically. Ask about drift, sync verification, licence terms or the privacy pipeline and an engineer replies.