# HOLON

> HOLON is a physical-AI hardware lab and engineering partner based in Hong Kong, Shenzhen and Dongguan. We source, validate and build data-capture hardware — egocentric rigs, trackers, grippers, wearables and edge-vision systems — with measured performance.

HOLON — physical-AI hardware lab. Hong Kong · Shenzhen · Dongguan. Contact: Atul Kamble (Founder), atul@holonai.ai, WhatsApp +852 6994 1597.

## Practices
- [Robot-learning data hardware](https://www.holonai.ai/practices/robot-learning-data): Egocentric headsets, body and object trackers, UMI grippers and synchronised multi-camera systems for teams training robot foundation models.
- [Wearables & health sensing](https://www.holonai.ai/practices/wearables-health): Connected collars, smart footwear, rings, earbuds and gloves — taken from a proven platform to a product you own, with your data on your cloud.
- [Edge vision & AI boxes](https://www.holonai.ai/practices/edge-vision): Rockchip, NVIDIA Jetson and Hailo edge boxes, smart cameras and GMSL camera systems for retail analytics and industrial inspection.
- [Data-collection operations](https://www.holonai.ai/practices/data-operations): Egocentric and manipulation data collection for robotics labs — 10,000+ hours a month across five Southeast Asian countries, in commercial and household environments.

## Research
- [Compute for Capture Devices and Edge Boxes: RK3576, RK3588, Allwinner A733, Raspberry Pi 5 and Jetson Orin](https://www.holonai.ai/research/capture-head-compute-rk3576-rk3588-jetson): For a recorder that only acquires, timestamps and stores, a low-cost Rockchip or Allwinner board is enough. For six raw cameras or on-device models, RK3588 or RK3576 (6 TOPS NPU) covers mainstream detection at low power; Jetson Orin is the choice when you need CUDA, larger models or the GMSL camera ecosystem. Benchmark your own pipeline, not TOPS.
- [Edge AI Boxes for Retail and Inspection: A Hardware Checklist That Survives the Pilot](https://www.holonai.ai/research/edge-ai-box-retail-analytics): Edge-vision pilots usually fail on hardware operations, not models: throttling in sealed enclosures, camera streams that drop frames, and no update or recovery path. Choose the box by measured FPS and latency for your model at maximum ambient temperature, keep identifying data on the device, and design remote updates and health telemetry before the first store.
- [Egocentric Data-Capture Hardware: 2026 Buyer's Benchmark](https://www.holonai.ai/research/egocentric-capture-hardware-2026): Egocentric capture hardware for robot learning falls into five classes: phone rigs, stereo RGB-IMU headsets, four-to-six-camera SLAM headsets, modular ego-plus-wrist-plus-gripper kits, and custom builds. Most teams should start from an existing SLAM-class headset and pay for targeted changes — frame rate, IMU rate, exposure timestamps, raw output — rather than a ground-up build.
- [Global-Shutter Sensors for Robot Tracking and Egocentric Capture: OV9281, AR0234, IMX296 and Alternatives](https://www.holonai.ai/research/global-shutter-sensors-robotics): For monochrome tracking cameras, the OV9281/OV9282 (1280×800, 3 µm, up to 120 fps) remain the default; OV2311 adds 2 MP with enhanced near-infrared. For egocentric RGB, AR0234 (1920×1200, 2.3 MP, global shutter) is the common choice, with IMX296 and OG02B10 as alternatives. The module, lens and trigger path matter more than the sensor.
- [The Hardware RFP Clauses That Decide Who Owns Your Product: Firmware, Data, Calibration and Design Files](https://www.holonai.ai/research/hardware-rfp-clauses-firmware-data-ownership): Before the first sample, a hardware RFP should require: all device data routable to a server you control; complete firmware source at delivery or in escrow; per-unit calibration files in an open format; a full BOM with manufacturer part numbers; design-file transfer terms; NRE priced separately from unit cost; and a measured acceptance test that gates payment.
- [Holon Labs: Event Cameras for Ultrafast Defect Inspection](https://www.holonai.ai/research/lab-event-camera-defect-inspection): Event cameras report per-pixel brightness changes asynchronously with microsecond-scale latency instead of full frames, so fast-moving or vibrating defects that blur in a frame camera stay sharp in event data. Combined with computational reconstruction, they make it practical to inspect dynamic processes at speeds and data rates frame cameras cannot match.
- [Holon Labs: Neuromorphic Imaging for Tracking Around Corners](https://www.holonai.ai/research/lab-neuromorphic-tracking-nlos): Non-line-of-sight tracking infers where a hidden object is by analysing light it scatters onto visible surfaces. Event cameras make this efficient because they record only the small changes in that scattered light, cutting data and latency compared with frame-based approaches.
- [MIPI, GMSL2, USB3 or 10GbE? Choosing the Camera Link for Multi-Camera Robots](https://www.holonai.ai/research/mipi-gmsl-usb3-10gbe-multicamera): Use MIPI CSI-2 when cameras sit within about 30 cm of the processor; GMSL2 or GMSL3 when they are up to 15 m away on a moving robot and you need power, control and trigger on one coax; USB3 for quick prototypes under about 3 Gbps total; and 10GbE when a wearable must send raw multi-camera streams to a backpack or base computer.
- [Can a Pet Collar Measure Heart and Breathing Rate? An Engineering Evidence Plan](https://www.holonai.ai/research/pet-collar-heart-respiration-sensing): A collar can plausibly estimate resting heart rate and respiratory rate from mechanical signals when the animal is still, with signal-quality gating and animal-specific validation. It cannot measure core temperature, blood oxygen or blood pressure on its own. Claims should be limited to what a synchronised reference has verified, and the device must report 'insufficient signal' instead of guessing.
- [How to Read a Shenzhen Camera-Module Quote (and the Eight Questions That Expose a Bad One)](https://www.holonai.ai/research/reading-shenzhen-camera-module-quotes): Treat every camera-module quote as a hypothesis. Confirm the exact sensor, the delivered resolution, frame rate and format per interface, whether a trigger pin is exposed, IMU part and rate, measured distortion for the actual lens, per-unit calibration, SDK status and what the price includes. In our 2026 reviews, most datasheets contradicted themselves on at least one of these.
- [Field Notes: Running Egocentric Data Collection Across Southeast Asia](https://www.holonai.ai/research/sea-egocentric-data-operations): Workplace egocentric collection succeeds or fails on operations: employer consent before the first session, trained participants, devices that are charged, clean and correctly mounted, protocol-specific QA that catches duplicates and filler footage, and reliable upload. Hardware choice matters, but process discipline decides yield.
- [The Shenzhen Factory-Visit Playbook for Hardware Founders](https://www.holonai.ai/research/shenzhen-factory-visit-playbook): A productive factory trip is decided before you land: pre-qualify suppliers against a written spec, send the agenda and questions in Chinese a week ahead, see samples and test equipment rather than showrooms, debrief each evening, and leave with a written comparison and next steps per supplier.
- [Timing Is the Spec: Synchronisation Requirements for Robot-Learning Capture Hardware](https://www.holonai.ai/research/timing-is-the-spec): For robot-learning capture, 'synchronised' must mean that every sample carries the time its light or motion was actually acquired, on one shared clock. Practical targets used by frontier teams are under 100 µs between tracking cameras, under 500 µs camera-to-IMU, and under 1 ms for RGB and radio events — verified by measurement, not vendor statements.
- [Body and Object Trackers for Robot-Learning Data: LED Pucks, Self-Tracking Pucks, IMU Suits and Custom Designs](https://www.holonai.ai/research/trackers-for-robot-learning): Trackers for robot-learning data use one of four architectures: IR-LED constellations seen by headset cameras, self-tracking camera pucks, IMU-only suits, or hybrids with UWB ranging. Off-the-shelf options are fine for testing, but most cap tracker count, compute pose on-device and lack raw, synchronised data — the requirements a training pipeline needs.
- [White-Label or Custom? A Dual-Track Strategy for Funded Wearable Startups](https://www.holonai.ai/research/white-label-vs-custom-wearables): Run both. A white-label device validates demand and the app in roughly 4–8 weeks; a custom design, typically 3–4 months to prototype, builds ownership and margin in parallel. Switch volume to the custom track once it passes validation. In both tracks, insist that device data can flow to your own cloud.

## Hardware Index
- [Hardware Index](https://www.holonai.ai/hardware): 60 capture devices, camera modules, trackers, camera links, compute boards and components sourced from a network of 100+ suppliers.

## Work
- [Moving an egocentric data company from phone rigs to an owned stereo capture device](https://www.holonai.ai/work/egocentric-data-company-device): A data company collecting workplace egocentric video on head-mounted smartphones needed a purpose-built, owned device. We ran a factory-side discovery, wrote the vendor specification with firmware-ownership clauses, and supported a low-cost stereo global-shutter recorder build.
- [Turning a frontier lab's egocentric-capture requirement into a buildable, testable hardware programme](https://www.holonai.ai/work/frontier-robotics-egocentric-program): A robot foundation-model team needed head-worn capture hardware with microsecond-class synchronisation, raw data and many body trackers. We translated the requirement into vendor-testable acceptance criteria, mapped the Chinese supply base, and structured an M1 fleet programme around an existing SLAM platform instead of a ground-up build.
- [Taking a connected pet-health collar from supplier dependency to an owned sensing platform](https://www.holonai.ai/work/pet-health-collar-owned-platform): A pet-health company shipping a white-label collar could not see or validate its own data. We audited the supplier device, then designed an owned research platform — sensing front end, temperature path, raw data, open manufacturing package — on a fixed four-week sprint.
- [10,000+ hours a month: egocentric data collection across Southeast Asia](https://www.holonai.ai/work/sea-data-collection-operations): We run egocentric and manipulation data collection across Malaysia, Indonesia, Vietnam, Thailand and the Philippines — more than 10,000 hours a month in commercial and household environments, on phones, egocentric glasses and tactile gloves, with capacity growing every month.
- [Smart footwear: re-engineering a connected shoe into a foot-mounted fitness instrument](https://www.holonai.ai/work/smart-footwear-impact-sensing): A high-volume footwear brand already selling a connected shoe wanted a smaller, longer-lasting, smarter next generation. HOLON took the brief apart — power budget, sensing, firmware state machine, packaging — re-scoped it into a buildable product, and led the sensing-module architecture, PCB form factor and firmware plan.

## Services
- Factory visit management (Duration of visit): Agenda, factory scheduling, Mandarin-led conversations, an embedded-systems specialist where needed, evening debriefs and a written recommendation.
- Spec & Vendor Sprint (2 weeks): Requirement → acceptance criteria → vendor landscape from our database → qualified shortlist → samples ordered. The fastest way to know what exists before you fund a custom build.
- Validation & Timing Lab (1–2 weeks): We measure a candidate device: exposure alignment, camera-to-IMU offset, frame drops over 30 minutes, raw-data access and calibration quality. Written report with raw data.
- M1 Build Programme (5–10 weeks): Fixed-scope prototype or small fleet with a qualified partner. Milestone payments tied to a live demo and a measured acceptance test.
- Production & supply (Ongoing): BOM procurement, QC oversight, certification coordination and logistics to your location, with full BOM and manufacturer part numbers.
- China Hardware Retainer (Monthly): An embedded hardware team on the ground: vendor management, factory visits, sampling, QC and escalation — reporting to your engineering lead.

## Team
- Atul Kamble, Founder
- Dr. Qianfeng Yin, Chief Scientist
- Priyank Patel, Hardware Engineering
- Dr. Shuo Zhu, Head, Vision & Timing Lab
- Qu Chao, Imaging Lead
- Amy Huang, Head of Supply Chain

## Optional
- [Glossary](https://www.holonai.ai/glossary)
- [Supplier network](https://www.holonai.ai/network)
- [Holon Labs](https://www.holonai.ai/lab)
- [About](https://www.holonai.ai/about)