FLOYD
--° SEOUL · LOADING

00 · ROVI, VISION EDGE DEVICE.

FLOYD

Seeing iswhere we start.

We hand the recognised pose to the robot as motion in its own frame. 2D, 2.5D and 3D run on one platform.

FLOYD — without Vision, your Robot is essentially blind.

DETAILS

FLOYD

Advanced Smart Vision Solution with AI

Optimised AI models make a wide range of image recognition applications easy to deploy. A high-performance EDGE DEVICE sharply lowers the cost of introducing and running AI systems, so vision runs efficiently on site.

KEY FIGURES · COMPANY DECK 2026

  • 3D BIN PICKING FROM A 2D CAMERA — World first · commercialised
  • BOP 6DOF POSE — TOP 10 · global benchmark, 2023
  • DEPLOYMENT COST · THROUGHPUT — 1/10 · 120% · vs. existing 3D systems
  • PHARMA INSPECTION VISION — 14 units · delivered, cumulative to 2025

CORE TECHNOLOGY

  • Image processing
  • Deep learning (AI)
  • Edge devices — Samsung LSI Exynos · Qualcomm Snapdragon · NVIDIA Jetson · RENESAS RZ/T

FIELDS

  • Object detection · tracking
  • 6DOF pose estimation
  • Grasping · gesture
  • 3D tracker · SLAM

BUSINESS

  • Robot vision system — ROVI, ROVIX
  • 2D / 3D vision system
  • AI inspection for secondary batteries
Ask about this solution →

01 · PICK · VLA · SHIPPING

ROVI2D / 2.5D

World first3D bin picking from a 2D camera, commercialised

DETAILS

ROVI2D / 2.5D

A vision system that works with industrial and collaborative robots

Mounted on industrial and collaborative robots, it estimates a part’s position and 3D pose from a 2D camera and hands them over in robot coordinates.

World first3D bin picking from a 2D camera, commercialised · 1/10 the cost and 120% throughput vs. the previous 3D system (Myoungshin Gunsan press line)

PROBLEM

  • When a robot picks stacked parts, the pull shifts and tilts the rest, so the line could not be automated
  • Existing 3D vision systems are expensive to introduce

APPROACH · 3D OBJECTS FROM A 2D CAMERA

  1. Select target part
  2. 2D pose (template matching)
  3. 3D pose from 2D-3D correspondence
  4. 3D pose optimisation
  5. Result to the robot

APPROACH · APPLICATIONS

  • Vision guided robot (VGR)
  • Object location / searching
  • Colour detection · histogram inspection
  • Presence / absence inspection
  • Geometric measurement

SET-UP · ROVI-2D

CameraSony 13MP RGB (rolling shutter) · 3A (AF/AE/AWB)
Resolution13MP 4160×3120 to 1.2MP 1280×960 · YUV2/NV12
FOV / AFH50° × V39° · 100 mm to ∞
DeviceQualcomm · Embedded Linux · 90 × 60 × 34 mm · 220 g
UXROVIX 1.0 · hand-eye / camera calibration · Ubuntu 18.04+ / Windows
Interface / powerTCP/IP, UDP · 24 VDC ± 20%
Protection / temp.IP66 · 0 to +60 ℃ (device)

PERFORMANCE · EVIDENCE

  • Commercial 3D bin picking of industrial parts from a 2D camera — provisionally assessed as a world first
  • Myoungshin Gunsan press line — 1/10 the cost of the previous 3D system, 120% overall throughput
  • Patents PCT/KR2021/006466 image recognition processing device · PCT/KR2021/006467 rotated image recognition method

CASES · MATURITY

  • Shipping — launched 2022, in production
  • Doosan Robotics — cobot OEM (Smart Vision Module), system upgraded onto the ROVI platform
  • Myoungshin — selected vision supplier for the e-CAR body M-module line
  • Processes — collaborative assembly, assembly fault inspection, planetary gears, engine blocks, PCB assembly, pick & place, vehicle shipping inspection

CASES · M-MODULE ROLL-OUT

  1. Gunsan (press / assembly / body)
  2. Yesan / Cheonbuk
  3. Other domestic
  4. Overseas (US, India …)
Ask about this solution →

02 · PICK · VLA · POC

ROVI3D

TOP 10BOP 6DOF pose global benchmark

DETAILS

ROVI3D

A vision system for accurate three-dimensional pose estimation

Estimates the 6DOF pose of overlapping, crossing 3D objects and lets the robot decide its own picking order.

TOP 10BOP 6DOF pose global benchmark · commercial methods (2023)

PROBLEM

  • Overlapping, crossing 3D parts are hard to detect, pose and order for picking
  • Collecting training images on the shop floor takes time and money

APPROACH · AI MODEL FLOW

  1. ROVI.net training images (PBR)
  2. FFB6D-PBR pose estimation
  3. Grasp learning (deep RL)
  4. 3D bin picking

APPROACH · HOW IT LEARNS

  • Physically based rendering (PBR) generates synthetic data for many spatial positions
  • Thousands to hundreds of thousands of images, generated to suit each environment
  • Learns occlusion cases in advance for a robust model

SET-UP

CameraTOF or RGB-D camera
DeviceHigh-performance barebone PC
UXROVIX 2.0 · hand-eye / camera calibration · Isaac (NVIDIA) / ROS / Unity
OSUbuntu 18.04+ / Windows

PERFORMANCE · EVIDENCE

  • BOP Challenge 6DOF pose estimation — top 10 among commercial methods (2023)

CASES · MATURITY

  • POC — development complete (2023)
  • POCs — Myoungshin (Gunsan), LG Innotek (Startup TechBlaze)
White paper Ask about this solution →

03 · PICK · VLA · ROADMAP

ROVIPolar

4 angles4 polarisation directions at once

DETAILS

ROVIPolar

Picking even the clear plastic parts that ToF and stereo miss — seeing them through polarisation

One polarisation camera captures RGB and polarisation (DoLP · AoLP) pixel-aligned, and recovers depth and 6D pose of transparent and shiny parts.

4 angles4 polarisation directions at once · 0° / 45° / 90° / 135°, pixel-aligned with RGB

PROBLEM

  • ToF, active stereo and structured light return holes or distorted depth on transparent (PET, acrylic, PC) and glossy surfaces
  • The 6D pose engine (FoundationPose) is documented not to work on transparent or glassy objects, so parts whose depth breaks cannot be picked with the existing pipeline

APPROACH · PIPELINE

  1. Raw polarisation image (12-bit)
  2. 4-angle split · Stokes
  3. RGB · DoLP · AoLP
  4. Physics-based normals · depth
  5. 6D pose (FoundationPose + ICP)

APPROACH · FEATURES

  • Polarisation read as physical quantities (refractive index, zenith, azimuth), not as extra image channels — a re-implementation of PolarDepth (ICML 2026)
  • Recovered depth feeds the existing AI ToolBox 6D pose pipeline unchanged — same data conventions and engine
  • Reflection-suppressed image (I_min) — brightness with glossy reflections filtered out

SET-UP

CameraLUCID Triton TRI050S1-QC (Sony IMX264MYR polarised colour · 5 MP · global shutter)
Polariser0° · 45° · 90° · 135° — per pixel, aligned with RGB
Lens · distance16 mm C-mount · 1 m working distance (528 × 442 mm field)
ComputeNVIDIA GPU PC (RTX 4080 class)

PERFORMANCE · TARGETS ARE DEVELOPMENT KPIS, PAPER VALUE FROM PUBLISHED POLARDEPTH (ICML 2026)

6D pose error≤ 5 mm / ≤ 3° @ 1 mTARGET
Depth δ1, transparent areas≥ 0.75TARGET
Speed≥ 5 fps (depth alone ≥ 20 fps)TARGET
Depth δ1, transparent (paper)0.859 — RGB-only depth model 0.547REFERENCE

CASES · MATURITY

  • Roadmap — capture and processing tools done, real-camera recording started; polarisation depth and pose integration in development (2026)
  • Target — picking injection-moulded clear plastic parts (PC, PMMA, PET), later inspection of transparent and shiny parts
White paper Ask about this solution →

04 · INSPECT · VL · COMMERCIALISING

ROVIInspection

OK / NGReal-time verdict

DETAILS

ROVIInspection

Product defect inspection

Finds anomalous regions first, classifies the defect and its location, then calls OK or NG.

OK / NGReal-time verdict · MES logging

PROBLEM

  • Surface defects such as scratches, black spots, burrs and gate burn must be judged and logged at line speed
  • Hard-to-judge defects in transparent plastics, transparent containers and plastic materials

APPROACH

  1. ROVI-SEG anomaly regions
  2. ROVI-DET defect class · location
  3. OK/NG verdict
  4. MES log

APPROACH · ON SCREEN (HEADLESS DEVICE + COMPANION APP)

  • ROI
  • Anomaly heatmap
  • Defect class
  • OK/NG result

SET-UP · RECOMMENDED

CameraTRI050S-MC high-res + lens
LightingCoaxial / polarised
Edge deviceGPU RTX3060 or better
OSWindows / Ubuntu 22.04

CASES · MATURITY

  • Commercialising — UX packaging and commercialisation (2026)
  • Transparent plastic, transparent container and plastic material defects — under review with manufacturers

CASES · SECONDARY BATTERY INSPECTION (X-RAY)

  • Innometry — inspection automation software for an e-CAR battery line (2022)
  • Trained on ~30,000 X-ray images · super-resolution for image quality
  • Anode and cathode edges and end points with CPM / OpenPose
Ask about this solution →

05 · COUNT · VL · COMMERCIALISING

ROVICount

14 unitsPharma inspection vision

DETAILS

ROVICount

Identifies a product and counts identical units automatically with 3D vision

Identifies a product and counts identical units with 3D vision, without a human hand.

14 unitsPharma inspection vision · delivered, cumulative to 2025

PROBLEM

  • Identical barcodes cause frequent miscounts of OTC drugs
  • Barcode reading takes time

APPROACH · PHARMA CHECK FLOW

  1. Order list · feeding
  2. AI vision count
  3. Order match
  4. Pack · ship

APPROACH · APPLICATIONS

  • Object counting from 3D depth
  • Product registration by barcode / QR
  • Per-product quantity control · real-time checking

SET-UP

CameraTOF camera (LUCID)
Edge devicenon-AI
UXHeadless device + companion app
OSWindows / Ubuntu 22.04

PERFORMANCE · EVIDENCE

  • Pharma detection solution — 14 units delivered in total as of 2025
  • 4 of Korea’s top 10 pharma distributors · JC Healthcare

CASES · MATURITY

  • Shipping — pharma detection solution in Eternity Solution’s ESR-P100 auto inspector (2024)
  • Commercialising — ROVI-Count UX packaging and commercialisation (2026)
Ask about this solution →

06 · INSPECT · VL · COMMERCIALISING

ROVISWIR

No trainingShort-wave IR camera

DETAILS

ROVISWIR

Transparent film packaging inspection

Detects transparent film wrap — hard to see by eye — with short-wave infrared.

No trainingShort-wave IR camera · film wrap presence

PROBLEM

  • A regular camera can barely tell whether transparent film wrap is there

APPROACH

  1. Set inspection area
  2. Detect film wrap
  3. Presence verdict

SET-UP

CameraLUCID SWIR camera (short-wave IR)
Edge devicenon-AI · no training needed
UXHeadless device + companion app
OSWindows / Ubuntu 22.04

CASES · MATURITY

  • Commercialising — UX packaging and commercialisation (2026)
  • Film-wrapped products such as cup noodles · via SI partners or direct
White paper Ask about this solution →

07 · MEASURE · VL · COMMERCIALISING

ROVISports

< 0.5 sProcessing

DETAILS

ROVISports

A smart golf launch monitor with premier-grade accuracy at a sensible price

Detects ball and club, tracks keypoints, and computes shot data.

< 0.5 sProcessing · vs ~1 s for SkyTrak+ / Mevo+

PROBLEM

  • Pro launch monitors (Trackman 4, Foresight GCQuad) cost $20,000 or more
  • $2,000-class units skip or estimate face angle and dynamic loft (Mevo+ none, SkyTrak+ D-plane estimate)

APPROACH · AI MODELS

  1. Ball / club detection
  2. Ball tracking
  3. Club keypoint tracking (driver / iron)
  4. Shot data

APPROACH · FUNCTIONS

  • Real-time ball and club data · spin · launch angle
  • Per-club detection · robust ball tracking
  • GSPro API · OEM customisation

SET-UP

CameraDual high-speed mono camera
Lighting850 nm IR
Edge deviceNVIDIA Jetson · Ubuntu 22.04
UXHeadless device + companion app

PERFORMANCE · VS SKYTRAK+ / MEVO+

Ball speed1.8%ON PAR WITH MEVO+
LA / LD0.6° / 0.5°ON PAR
Total spin5–6%ON PAR WITH MEVO+
Spin axis3–4°ON PAR OR BETTER
Back / side spin450–550 / 350–450SLIGHTLY BEHIND
Face angle< 2°CLEAR ADVANTAGE
Attack / path< 2° / < 3°ON PAR
Dynamic loftMeasuredCLEAR ADVANTAGE
Processing< 0.5 sFASTEST

CASES · MATURITY

  • Commercialising — technology complete, product enclosure design under way (2026)
  • For launch-monitor distributors · OEM manufacturers
User manual Ask about this solution →

08 · MEASURE · VL · POC

ROVIEVS

10,000 fpsVirtual frame rate

DETAILS

ROVIEVS

A camera that sees only motion — tracks moving objects without blur, even in fast or dark scenes

An event camera that outputs only the pixels that change finds moving people and vehicles and reports ID, box and speed in real time.

10,000 fpsVirtual frame rate · events rebuilt into 0.1 ms frames, super slow motion against a 30 fps camera

PROBLEM

  • Frame cameras blur fast motion over the exposure time and lose targets against backlight or in low light
  • They process the still background in every frame, so data and compute are heavy, and finding moving objects needs a background model or a trained detector

APPROACH · PIPELINE

  1. Event input
  2. Noise filter
  3. Moving-object clusters
  4. Tracking (ID · box)
  5. Velocity (optical flow)

APPROACH · FEATURES

  • Events come only from where something moved, so no background removal is needed
  • Per-pixel µs timestamps — no exposure, no motion blur
  • Moving objects, IDs and speed without training data; AI detection adds person and vehicle classes
  • Own algorithms without the Prophesee commercial SDK — no software royalty per unit

SET-UP

CameraLUCID Triton2 EVS (Sony IMX636 · 1280×720 · 2.5GigE)
Lens6.5 mm C-mount
Computex86-64 PC · real time on CPU (NVIDIA GPU for AI detection)
OSUbuntu 22.04

PERFORMANCE · MEASURED ON THE DEMO, SPEC FROM THE SENSOR DATASHEET

Virtual frame rate10,000 fps (0.1 ms apart)MEASURED
Tracking update100 Hz (10 ms)MEASURED
Tracking timep95 3 msMEASURED
Live throughput9.8 Mev/s · no lossMEASURED
Dynamic range> 120 dBSPEC
Pixel latency≤ 100 µs @ 1000 lxSPEC

CASES · MATURITY

  • POC — real-time pedestrian and vehicle tracking with a Triton2 camera (2026)
  • Under review — counting fast-falling parts (pharma), tracking movers against backlight or in low light
User manual White paper Ask about this solution →

09 · UNIFY · PLATFORM · COMMERCIALISING

AIToolBox

DETAILS

AIToolBox

One vision platform for ROVI-2D, 2.5D and 3D

Three tiers: it trains in the cloud and decides at the edge. Vision solutions once split by use case become one platform that works with collaborative robots.

PROBLEM

  • 78% of robot makers outsource their vision AI
  • Solutions are fragmented by use case — 2D, 2.5D and 3D

APPROACH · THREE TIERS

CLOUD · ROVI.netA 3D model is enough — PBR rendering produces the training set.
AI CORE · SEG / DET / POSE2D, 2.5D and 3D share one pipeline.
EDGE · ROVI Device · ROVIXInference runs on site and logs to MES.

APPROACH · ROVIX ON-SITE SOFTWARE

CALIBRATIONAuto-matches the robot and vision coordinate systems on the worktable · distortion correction · Robot-Eye (fixed camera) / Hand-Eye (robot-mounted camera)
TEACHINGSingle / multi-target teaching · contour information and a unique ID · measurement · histogram masking · pixel counting for colour detection
DETECTINGMatches the target through scale, rotation and translation to detect and position it
HISTORYCollects vision processing data to show performance and progress, and to trace problems quickly

APPROACH · SUPPORTS

  • Hand-eye / camera calibration
  • Isaac (NVIDIA) / ROS / Unity
  • Vision simulation (digital twin · Isaac Sim)

SET-UP

CameraTOF or RGB-D camera
DeviceHigh-performance barebone PC (NVIDIA Jetson)
RobotUR, Doosan / Hanwha / Rainbow Robotics, etc.
OSUbuntu 22.04+ · ROS

SET-UP · SENSOR MATRIX — PICK THE SENSOR FOR THE TASK, KEEP ONE SOFTWARE

2D RGBROVI-2D / 2.5D · pharma detection — position and 3D pose from a 2D camera
TOF / RGB-DROVI-3D · ROVI-Count · AI ToolBox — depth-based pose and counting
High-res industrialROVI-Inspection — fine surface defects
SWIRROVI-SWIR — transparent film wrap
Polarised colour (Polarsens)ROVI-Polar — depth and pose of clear and shiny parts (in development)
High-speed mono × 2 + 850 nm IRROVI-Sports — fast ball and club tracking
Event camera (EVS)ROVI-EVS — tracking fast movers, also in low light
X-ray (customer equipment)Secondary battery inspection — anode and cathode edges
LiDAR / TOF 3DK9-K10 auto-docking — 6D pose at 1 m

PERFORMANCE · K9-K10 AUTO-DOCKING

  • 6D pose from the square receptacle’s geometry — ±5 mm / ±1° at 1 m · 15 FPS or more
  • LiDAR or TOF 3D sensor · physics-based docking simulation

CASES · LEAD POC FLOW

  1. Detect K9
  2. 3D pose
  3. Coordinate transform
  4. 3-axis alignment

CASES · MATURITY

  • Commercialising — commercial development (2026)
  • POC — K9-K10 auto-docking, due August 2026
  • Roadmap — UR+ URCap certification, listing targeted for H1 2027
User manual Ask about this solution →

10 · PICK · VLA · POC

AI ToolboxVLA

≤ 0.5 sFrom command to object choice

Unsure or out of schema → ask back instead of acting 5 Result reported in text Text command “Pick the blue square and move it to A” 한 / 영 · free wording 1 Command Parser LLM → ACTION JSON Schema gate · closed tokens 2 VLM Grounding Grounding DINO → box SAM → mask CV check · colour / shape / size 3 Pose 3D position + axis 6-DoF · FoundationPose (CAD) via calibration chain 4 Robot Executor approach → grasp → lift → move → place Isaac Sim · ROS · UR RTDE · safety volume RGB-D camera Next · WAM policy (Cosmos)
DETAILS

AI ToolboxVLA

“Move the blue square to point A” — pick and place in one sentence

A customer types an instruction and the robot finds, picks and moves the object — the first step towards AI Toolbox’s goal of a robot you use with text.

≤ 0.5 sFrom command to object choice · 0.2–0.5 s measured on the resident server

PROBLEM

  • Chat-style VLMs give inaccurate, unstable object coordinates (boxes); passed straight to a robot, they can make it pick the wrong object
  • End-to-end VLAs understand meaning well but generalise poorly on how a scene will physically change (NVIDIA technical blog)

APPROACH · PIPELINE

  1. Text command
  2. ① Parse (LLM → ACTION JSON)
  3. Schema gate
  4. ② Find the object (Grounding DINO + SAM)
  5. ③ 3D position · 6-DoF pose
  6. ④ Robot executes
  7. ⑤ Report

APPROACH · FEATURES

  • Modular VLM — the VLM only understands the command and finds the object; the proven control stack (Isaac Sim, ROS, UR RTDE) does the motion
  • Schema gate — Korean, English and synonyms normalised to closed tokens; anything outside the schema is asked back, not executed
  • VLM results cross-checked with colour, shape and size CV checks — when unsure, the robot asks instead of picking
  • Objects with CAD get a full 6-DoF pose from FoundationPose — assets proven in AutoDocking

SET-UP

Command parserLLM (Claude API or local Qwen2.5-7B) + pydantic schema check
Object finderGrounding DINO (open-vocabulary) + MobileSAM
Pose3D position + axis / FoundationPose 6-DoF (CAD objects)
ExecutionNVIDIA Isaac Sim · ROS · UR RTDE
CameraCalibrated RGB-D (Helios + RGB)

PERFORMANCE · MEASURED ON SYNTHETIC SCENES (2026-08), TARGETS ARE FIRST-DEMO KPIS

Response (resident server)0.2–0.5 sMEASURED
Target object position±4 px from ground truth (zero-shot)MEASURED
Colour trap (blue circle)Rejected by the shape checkMEASURED
Wrong-object picks0TARGET
Pick successsim ≥ 95% / real ≥ 90%TARGET
Command → placed≤ 30 s (recognition ≤ 3 s)TARGET

CASES · MATURITY

  • POC — VLM server and object-selection gate chain built and verified; Isaac Sim end-to-end, then the real UR (2026)
  • Next — evaluate a WAM (World Action Model, NVIDIA Cosmos) policy as the executor; multiple objects and chained commands
Ask about this solution →

11 · FOUNDING → PLATFORM

FLOYDRecord

  • 01Founding and first proof
  • 02Into customer lines
  • 03The product line completes
  • 04The platform turn
DETAILS

FLOYDRecord

From a product company to a platform company.

Based on the history in the company deck 2026 r2.

01 · Founding and first proof · FOUNDING

  • 2020 · FLOYD inc. founded · corporate research institute
  • 2020 · OEM production and delivery for Doosan Robotics
  • 2020 · Start UP NEST cohort 8 (Korea Credit Guarantee Fund) · TIPS startup selection and investment
  • 2021 · TIPS commercialisation grant · seed round · venture certification
  • 2021 · G-Valley startup · machine-vision MOU with NextEye

02 · Into customer lines · DEPLOYMENT

  • 2021 · Myoungshin e-CAR body line automation POC passed
  • 2022 · Myoungshin selects FLOYD as vision supplier for the M-module line
  • 2022 · Innometry: inspection automation software for an e-CAR battery line
  • 2023 · Hyundai Mobis: occupant pose estimation from an AI-3D camera
  • 2023 · RENESAS e-AI IDH & Resale Partner · AI voucher supplier

03 · The product line completes · PRODUCT

  • 2020 · ROVI-2D · 2021 · ROVI-2.5D · 2023 · ROVI-3D complete
  • 2024– · ROVI-INS, pharma vision inspection and the golf launch monitor complete
  • 2023 · COMEUP Stars 2023 The Stars of Stars · LG Innotek Startup TechBlaze
  • 2023 · STAGE Cheonan award · Booster IR bronze · EdgeUp:Cre8tors selection and investment

04 · The platform turn · PLATFORM

  • 2024– · AI Agent adopted · every vision solution built and produced AI-first
  • 2024– · AI Vision Total Solution — AI ToolBox in development
  • 2026 · K9 / K10 auto-docking POC — due August
  • 2027 · UR+ URCap certification — listing targeted for H1
Ask about this solution →

12 · CONTACT

Contact

DETAILS

Contact

Find out whether it fits your line first.

Tell us the task (pick, inspect, count, measure), send images of the part and your cycle time. We reply with feasibility and a recommended configuration, and can quote the SI, camera and robot in one go.

REVIEW REQUEST

    Name (required)

    Email (required)

    Subject

    Message

    Attachment (optional, up to 10 MB)

    One jpg · png · pdf · doc · xls · ppt · dwg · dxf · step file. Files are used only to review your request.

    File link (optional, share large files or videos via Google Drive or similar)

    CONTACT

    MAILFLOYD@FLOYD.KR
    TEL070-8676-7976
    ADDRRm. 407, Gasan House D Wise Tower, 75-15 Digital 1-ro, Geumcheon-gu, Seoul, South Korea (08592)
    00 / 12