TB-2972 · REV B · Technical newsheet

Machine Vision & InspectionDevice profile

99.5 Percent Pick Accuracy: Eureka Robotics Details Bin-Picking at Scale

Eureka Robotics reports 30 million production picks at Toyota and Pratt & Whitney at 99.5 percent accuracy, with edge AI controllers and a supermodel trained on millions of parts.

By Grace Kim4 min read708 words

Features

  • Eureka Robotics reports 30+ million production picks at Toyota and Pratt & Whitney with 99.5 percent accuracy.
  • Target cycle times run two to three seconds per pick; on-site software upgrades take about one hour.
  • Company spun out of Nanyang Technological University, Singapore, in 2018; partners with Shelley Automation for North America.
  • Supermodel pre-trained on millions of industrial parts and synthetic images enables zero-shot picking.
  • Edge-based containerized controllers keep all images and CAD files on the factory network with no cloud dependency.
Deploying Vision-Guided Robotics in Real-World Factories
Device photoDeploying Vision-Guided Robotics in Real-World Factories — kalhh / pixabay

Eureka Robotics has completed more than 30 million picks in production at clients including Toyota and Pratt & Whitney, with picking accuracy at 99.5 percent and cycle times of two to three seconds. Francisco Suarez of Eureka Robotics laid out the figures and the engineering behind them in an interview with Sharon Spielman, head of content at VSD, on the Visions machine vision podcast.

The company spun out of Nanyang Technological University, Singapore, in 2018. It builds physical-AI vision systems for high-mix bin picking and precision handling across automotive, aerospace and electronics. A strategic partnership with Shelley Automation aims to extend deployments across North American factories: Shelley supplies regional systems integration, and Eureka supplies hardware-agnostic AI controllers and vision software. Suarez said the collaboration targets a reduction of deployment timelines "from maybe months to days."

What replaces rule-based geometric matching?

Suarez framed the core problem with traditional vision: rigid rule-based geometric matching breaks under lighting shifts, part reflections, and small part variations, and demands manual tuning for every part.

Eureka's answer is a "supermodel" — a foundational neural network pre-trained on millions of industrial parts plus millions of synthetic images. The model generalizes across surfaces, occlusions and random orientations, and enables zero-shot picking for known items. Synthetic training is domain-specific but randomized, which Suarez said makes the models effectively invariant to lighting shifts — a lesson, he noted, learned the hard way with factories that run in uncontrolled daylight.

For dark rubber and plastic parts that absorb light, Eureka often applies structured light to improve 3D sensing.

How does the edge controller handle updates and data?

Everything runs on the controller — edge computing, in containerized form. Tuning, configuration changes and upgrades are developed offline with synthetic data and, when available, real captured data. Deployment to production then takes about one hour, Suarez said.

The open platform supports robots from different brands and is arm-agnostic. Running locally gives the customer control over data, consistent low latency, and zero cloud dependency. "You are not sharing part geometries or any proprietary information," Suarez said — nothing leaves the factory network, and the controller needs no external connection.

Why does sensor quality decide bin-picking success?

The 3D camera supplies point clouds from which the system computes surface normals, essential for estimating part orientation and planning collision-free motions into the bin.

Sensor quality directly determines point-cloud quality. Shiny, oily or black metallic parts produce what Suarez called phantom surfaces — ripples in the point cloud. Reflection suppression and resolution are therefore critical specifications.

What did 30 million picks teach?

Two metrics define maturity for customers like Toyota and Pratt & Whitney, according to Suarez: proving 99.5 percent picking accuracy, and meeting two-to-three-second cycle times, where the vision system must not become the bottleneck.

Maturity itself, he said, is defined by edge-case handling: damaged bins, overlapping parts, clutter, oily and shiny parts. These should never trigger line stops, so edge cases belong in testing early rather than on the happy path.

His deployment best practices:

  • Fix mechanical design first: tooling clearance, gripper design and reachability cause more pick failures than vision miss-detections.
  • Benchmark with dirty, scratched and reflective parts, not pristine samples.
  • Place cameras to minimize shadows and bin occlusion.

Eureka publishes these practices in a field guide for engineers on its website.

The software interface grew iteratively from internal low-level tooling into wizards that import CAD files and step through calibration, pick-point definition and workspace setup, so operators at any skill level can configure the system.

On safety, Suarez cited deterministic grasp calculation, deterministic motion planning, and hardware interlocks for both system and operators. On security, on-premise edge processing keeps images and CAD files inside the factory.

Looking ahead, he named two trends: simulation acceleration, where entire cells train and stress-test overnight in physics-engine digital twins to cut on-site commissioning, and multimodal action models that move from passive part detection to end-to-end see-and-act policies adjusting dynamically mid-motion.

The open question for integrators: whether simulation-first commissioning can hold the promised 99.5 percent accuracy on the first day of production, not just after months of on-site tuning.

via Vision Systems Design (Source)

Filed under

  • bin-picking
  • industrial-ai
  • 3d-machine-vision
  • edge-computing
  • robotic-handling
Share this article:

More from Grace Kim

Grace Kim

Show full bio

Correspondent covering companies, deals and strategy in Instruments, test and measurement at Testbench Report.

116 articles

Application notes

« Previous articleNext article »