All case studies
Robotics & AutomationENGINEERING STUDY

Vision-Guided Robotic Bin-Picking Cell with 3D Pose Estimation

In this robotics vision study, YantriX evaluated a vision-guided robotic manipulation cell—combining deep learning segmentation on Jetson Orin Nano, point-cloud pose estimation, and MoveIt 2 motion planning for picking unstructured parts from deep containers.

By Tirth Borda · Co-Founder · Robotics Systems Lead2 min readFactory Automation & Material Handling
Vision-guided robotic arm bin-picking workcell

Overview

Engineering Scope & Context

Robotics engineering study demonstrating unstructured bin-picking using YOLOv11-Seg instance segmentation, 3D surface normal extraction, and collision-free grasp trajectory planning.

Engagement: ENGINEERING STUDY

Discipline: Robotics & Automation

Project Type: Robotics Vision (Edge AI + 3D Pose Estimation + Motion Planning Study)

Application: Factory Automation & Material Handling

YantriX Scope Ownership

  • RGB-D camera calibration and eye-in-hand / eye-to-hand transformation mapping
  • Model training and TensorRT optimization for instance segmentation on Jetson edge compute
  • Point-cloud filtering, centroid estimation, and surface normal grasp vector calculation
  • MoveIt 2 motion planning workspace configuration and collision-matrix tuning
  • Grasp quality scoring and container collision boundary avoidance algorithms
Discuss a Robotics ProjectBack to all case studies

Key specifications

Core engineering parameters & stack.

TensorRT
On-Device Inference Pipeline
3D Pose
Bin-Picking Perception Study
0
External Cloud Dependencies
MoveIt 2
Trajectory Planning Integration

Objectives

What the project needed to achieve

  • Deploy real-time instance segmentation on Jetson Orin Nano with on-device TensorRT inference
  • Generate 6-DOF grasp pose vectors from segmented depth point clouds
  • Integrate perception outputs directly with MoveIt 2 trajectory planner
  • Achieve collision-free robot arm trajectory execution into deep storage bins

Challenge

Engineering constraint

Random bin picking remains a difficult automation challenge due to overlapping parts, reflective metallic surfaces, bin lip collisions, and variable ambient lighting. Cloud-dependent vision processing introduces unacceptable latency and network reliability risks. This study explored an entirely edge-deployed perception and grasp planning pipeline operating on embedded hardware.

Approach

How YantriX approached the work

  1. 01

    Trained a lightweight segmentation model on labeled industrial parts, converting weights to TensorRT INT8.

  2. 02

    Processed depth streams to generate localized surface normals and identify top-ranked grasp candidates.

  3. 03

    Connected the vision perception node to ROS 2 MoveIt 2 via custom service interfaces.

  4. 04

    Implemented pre-grasp approach and retreat offsets to prevent gripper interference with bin walls.

Outcomes

What improved by the end

  • Evaluated low-latency on-device TensorRT inference pipeline on embedded Jetson Orin Nano hardware
  • Demonstrated reliable 6-DOF grasp pose extraction across overlapping test components
  • Executed collision-free reach and retrieve trajectories in MoveIt 2 simulation environment
  • Eliminated all external cloud dependencies for privacy and latency control

Deliverables

Engineering Deliverables & Outputs

  • ROS 2 vision perception packages and launch scripts
  • Trained segmentation model weights and TensorRT deployment configuration
  • MoveIt 2 workspace configuration files and collision environment model
  • Technical evaluation and latency notes

Tools used

Stack and tooling

  • NVIDIA Jetson Orin Nano
  • YOLOv11-Seg
  • ROS 2 Humble
  • MoveIt 2
  • OpenCV

Impact

Business-level effect

  • Demonstrates a deployable edge vision architecture for integrating modern deep learning with industrial robotic manipulation

Conclusion

Robotic bin picking succeeds when vision perception and motion planning are tightly co-designed for edge execution rather than treated as isolated black boxes.

Working on a similar engineering problem?

Need to automate a part handling, inspection, or robotic manipulation workcell? Discuss your vision and kinematics requirements with our robotics team.

Discuss a Robotics Project

Tagged

  • Robotics
  • Computer Vision
  • Bin Picking
  • MoveIt 2
  • ROS 2
  • Jetson

Related case studies

Adjacent proof you can read next.

Continue exploring

Related blogs, services, and capability pages

Explore the engineering service and practical guides related to this work.

Service pages

From the blog

Similar Engineering Need?

Need help developing something similar?

Discuss your vision-guided robotic bin-picking cell with 3d pose estimation requirements, design challenges, or target specifications with our engineering team. NDA support is available where required.

Discuss a Robotics ProjectView Engineering Work→