Handforge Robotics records skilled people doing physical work in the buildings where they do it — and independently measures what a robot trained on that work can actually achieve.
What we are
We operate our own capture environments — hotels, co-living buildings and gyms across Metro Manila, thirty of them owned and operated and nine run by franchise partners — and we staff them with the people who already do that work for a living.
Two things come out of that. The first is a corpus of manipulation data with documented provenance: who performed the task, how long they have done it, in what environment, and whether the attempt met a written acceptance criterion. The second is something the industry does not currently have — a reproducible, independent measure of whether a robot policy can perform a task to specification.
We are not an annotation vendor and we do not build robots. We supply the evidence that sits between them.
The record
Capture frame · task 0042 · towel fold
Trajectory, object segmentation and provenance are shipped with the frame, not inferred from it.
One attempt at one task, by one named operator, in one instrumented environment — recorded from four synchronised viewpoints and delivered with the provenance intact.
The distinction matters commercially. A model can learn from a demonstration; a buyer can only defend a purchase that is documented. Every field below travels with every episode we ship.
| Visual | 4 × RGB-D viewpoints, 30 fps, hardware-synchronised to under 2 ms drift |
|---|---|
| Depth | Stereo pair plus active depth; per-frame point cloud retained |
| Kinematics | 6-DoF hand pose and 21-point hand skeleton at 60 Hz |
| Contact | Instrumented gloves where the task warrants: per-fingertip normal force |
| Segmentation | Per-frame masks for task-relevant objects only — no background labelling |
| Provenance | Operator identifier, role, years in role, site code, shift, consent record |
| Outcome | Pass or fail against a written acceptance criterion, plus failure class where applicable |
| Formats | LeRobot, RLDS or HDF5; MP4 media with Parquet sidecars |
| Licence | Non-exclusive by default. Exclusive windows of 12, 18 or 24 months by agreement |
Capabilities
Each runs on the same capture pipeline and the same operator pool. Most customers start with a corpus licence and move to a commissioned programme once their task list is settled.
Corpus
Curated manipulation corpora by environment class, released in versioned drops. Sample first, license after.
Programme
A dedicated capture programme against your task list, embodiment and acceptance criteria. Scoped in a written statement of work with an agreed rejection protocol.
Operations
Trained operators running deployed fleets in the loop, on shift, to a service level. Takeover latency and coverage are contracted, not estimated.
Evaluation
Human-scored evaluation of a policy against real tasks in environments the vendor has never trained on. Delivered as a signed report with the full trial record attached.
Independence
Every certification regime in robotics today addresses safety. UL 3300, the TÜV schemes and ISO 10218:2025 all ask whether a machine can injure someone. None of them asks whether it can complete the task it was sold to perform.
The consequence lands furthest downstream. Insurers writing performance cover on AI systems price that risk using performance data supplied by the vendor they are underwriting, because no independent source exists. Enterprise buyers comparing two robot vendors have no number either.
That gap is why we operate our own environments rather than renting them.
Every benchmark in machine learning eventually fails the same way: the test set finds its way into training data, scores inflate, and the number stops meaning anything. In physical AI the test set is not a file. It is a room, and a room can be controlled.
A subset of our sites is permanently ring-fenced for evaluation. Those sites are never captured for training data, never licensed, and never included in any corpus we sell. They are re-randomised between runs — linen types, clutter, lighting and layout — so that a policy cannot be tuned to a fixed arrangement.
We sell training data and we also evaluate policies. Those two activities cannot sit in the same entity without the conflict being managed explicitly, so it is:
| UL 3300 | Service, communication and personal-care robots. Electrical and mechanical safety. |
|---|---|
| ISO 10218:2025 | Industrial robot safety requirements. Learned behaviour is out of scope. |
| TÜV SÜD & Rheinland | Functional safety and conformity assessment. |
| Task performance | No scheme exists. A buyer’s only source today is the vendor’s own report. |
Transfer
The first question every robotics team asks is whether data recorded from a human hand is worth anything to a two-finger gripper. It is a fair question, and the answer is specific rather than promotional.
Human demonstration does not replace on-robot data. What it does is reduce how much on-robot data you need, and recent work puts a number on it: co-training on human video produced a 29.7% absolute gain in task success in the low-robot-data regime. The same work found the variable that decides whether transfer happens at all is the quality of the hand pose — which is precisely what separates a calibrated multi-camera capture from a scraped video.
Layer 1 · embodiment-independent
Per-frame masks and 6-DoF pose for every task-relevant object. A towel going from flat to folded in four steps is the same event whoever performs it — this layer carries no morphology at all and transfers to any embodiment without modification.
Layer 2 · retargetable
6-DoF wrist pose and a 21-point hand skeleton at 60 Hz, in the task frame. For a parallel gripper, aperture derives from thumb-to-index distance and approach from the wrist frame. For a multi-finger hand, the skeleton retargets joint-to-joint. You receive both the derived end-effector trajectory and the raw skeleton, so you can retarget with your own method rather than inherit ours.
Four synchronised RGB-D views, hardware-locked under 2 ms — not monocular estimation.
Layer 3 · the residual gap
A motion gap survives retargeting: human arms and robot arms move differently even when the hand pose is right, and closing it needs data from your hardware. That is what our teleoperation line is for — leader-follower capture on your embodiment, in the same environments, against the same task list.
Layer 4 · the verdict
Every episode carries a pass or fail against a written acceptance criterion, plus a failure class. That turns the corpus into something you can filter and weight by outcome, rather than a pile of footage in which successes and failures look identical.
We do not claim human demonstration solves the embodiment problem. It moves the bulk of the data requirement off your robots and onto people who already do the work — and then we capture the remainder on your hardware.
Co-training gain and the hand-pose finding: “What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?”, arXiv 2606.06627. Retargeting approaches vary by target morphology; we deliver the inputs rather than prescribe the method.
Delivery specification
Everything a delivered drop contains, published in full. Check it against your ingestion pipeline; if a field you need is missing, say so before we scope anything.
Sample drops carry the identical structure at reduced volume, with the provenance record intact.
Delivered by S3 or GCS transfer with per-file checksums in the manifest. LeRobot and RLDS conversions ship alongside the source parquet rather than replacing it, so you can ingest either. HDF5 on request.
Nothing in a drop requires our tooling to read: MP4, Parquet and JSON throughout.
{
"episode_id": "ep_0000042",
"task": {
"id": "TOWEL_FOLD",
"taxonomy": "handforge/hospitality/v3",
"criterion": "four-fold, edges aligned <2cm"
},
"outcome": {
"accepted": true,
"failure_class": null,
"scored_by": "reviewer_014"
},
"operator": {
"id": "op_021",
"role": "housekeeping supervisor",
"years_in_role": 11,
"consent_ref": "c_2027_0113"
},
"environment": {
"site": "MNL_SITE_03",
"class": "hotel_guest_room",
"held_out": false
},
"capture": {
"cameras": 4,
"fps": 30,
"sync_drift_ms": 1.4,
"duration_s": 8.42,
"hand_hz": 60,
"force": false
},
"licence": "non-exclusive-perpetual"
}
Field values above illustrate the structure of a delivered record. Handforge Robotics is in formation; capture programmes open in 2027 and sample drops are issued against a signed evaluation licence.
Environments
Our focus is deformable-object manipulation in confined service environments — linen, towels, garments, packaging and bedding handled in small, cluttered, human-scale rooms.
It is the manipulation class where simulation transfers worst, so a real demonstration is worth most. The capture pipeline itself is general, and we scope programmes outside this focus on request.
Primary focus
Bed strip and make, towel fold, turndown, linen handling. Eighteen properties under own management.
The tightest space a service robot will meet: hard surfaces, reflections, wet tile and no room to manoeuvre.
Sorting, shaking out, folding and stacking at volume. The hardest deformable geometry in the set.
Prep, plating, pass, dish return and pantry restock — including wet, high-volume work at the dish station.
Equipment reset, re-racking, floor cleaning, towel service and stockroom flow. Twenty-one sites.
Induction, singulation, tote handling and exception resolution on live distribution floors.
Additional programmes — scoped to a customer brief
Fabric handled to millimetre tolerance at speed — the industrial end of the same deformable problem.
Irregular, fragile, no two items alike, and damage is immediate. The extreme case of object variance.
Facing, replenishment, case handling and stockroom flow in grocery and convenience formats.
Adhesive, temperature-dependent and irreversible. Currently unsolved, and captured on request.
Tool use, heavy parts and precise fitting — a manipulation class the service verticals never exercise.
Precision assembly, field maintenance, care environments. If it is physical work done by professionals, we can scope it.
Operations
Capture happens in two places. On live sites, operators work their normal shift wearing a lightweight harness that records without changing how the job is done. On our capture floor, the same environments are rebuilt as instrumented bays where lighting, layout and object set can be controlled and repeated.
Teleoperation runs from the same floor, staffed on a follow-the-sun rota out of UTC+8 — which covers a European morning and an American evening from a single shift pattern.
Capture floor · Makati
Teleoperation station
Instrumented bay
Episode review
Facility imagery is illustrative of the operating model. Handforge Robotics is in formation and the capture floor is under development.
People
Our operators are recruited from the sites where they already work. They are not taught the task in order to record it — they have been doing it for years, and that is the asset.
The Philippines maintains a national competency framework covering roughly 320 occupational qualifications, decomposed into individually assessable units and externally audited — the same system that keeps the country on the IMO White List for seafarer certification. That gives us a defensible way to state what an operator is qualified to do, rather than asserting it.
Founders
Co-founder & Chief Executive
Mark has spent more than a decade building and operating physical service businesses in Metro Manila. He founded and runs PULS — Philippines Urban Living Solutions — which operates the MyTown co-living brand across fourteen buildings and the MyStay hotel brand across four, all in Makati. He also operates GoGym, with twelve owned locations and nine franchise partners.
Those sites are our capture estate. They employ housekeeping, laundry, maintenance and floor staff performing exactly the manipulation tasks that service robotics is trying to learn, in exactly the confined environments those robots will be deployed into.
Before PULS he established a business process outsourcing company in the Philippines, which is where our operating model comes from: recruiting, training and running shift-based service delivery to a contracted quality standard for enterprise customers abroad. He is also a general partner at Singularity Dynamics, a Hong Kong-based deep technology fund, which is how he came to the physical-AI problem.
Co-founder & Chief Operating Officer
Janna is the co-founder and chief executive of Bridge Access, a Philippine financial technology company providing salary-linked lending, automated savings and salary advances to lower-income workers, repaid through payroll deduction. It sits within Bridge Southeast Asia, whose stated purpose is to invest in the future of work.
Bridge Access is built on payroll integration with Philippine employers, so she has spent years working inside the systems that pay, verify and retain the same workforce we record — hourly service staff, on shift, at scale.
At Handforge she runs operations: the capture floor, the operator base, and the standards above that make an episode defensible.
We are hiring capture leads, QA specialists, teleoperation staff and pipeline engineers in Metro Manila. Domain experience on the floor counts for more than a CV.
Data protection
Human video of workplaces, facility layouts and robot telemetry carry personal and sensitive-data implications in every market we sell into. We assume the review and arrive with the paperwork.
| European Union | GDPR processor terms; EU AI Act training, validation and test-data governance. Data processing addendum, subprocessor list, retention schedule, EU storage region on request. |
|---|---|
| Japan | Cross-border transfer mechanisms with clear notice, named purpose and vendor accountability. CBPR-aligned terms; purpose stated per statement of work. |
| China | Personal, sensitive and “important data” exposure in video, site maps and telemetry. China-local processing only, no offshore raw data, counsel engaged per programme. |
| Philippines | Data Privacy Act of 2012 and National Privacy Commission obligations. Data protection officer appointed, breach notification terms, contractual deletion certificates. |
| All markets | Two-person approval on every export, least-privilege access, full access logging, and no reuse of one customer's commissioned data for another. |
Contact
Tell us the embodiment, the task list and the environment class. We come back with a sample delivery and a written acceptance criterion before any commercial conversation.
Commercial
Everything on this site routes through one inbox. Use the form below, or write to us directly and say which of the four you are asking about.
hello@handforgerobotics.com →Careers
Capture leads, QA specialists, teleoperation staff and pipeline engineers, in Metro Manila. Tell us what you have run on a floor.
careers@handforgerobotics.com →Tell us what you are trying to build and we will come back with a sample delivery, a written acceptance criterion and an indicative schedule. Nothing on this form commits you to anything.
Your enquiry is ready to send. We have opened your email client with everything filled in. Press send there and it reaches us.
Nothing happened? Copy the details from this link and email them to hello@handforgerobotics.com. We reply within two business days, Manila time.