Semantic Fetch Robot

A ROS 2 mobile manipulation system that accepts a natural-language object request, navigates an indoor environment, locates the target using open-vocabulary vision, and physically retrieves it using a mounted robotic arm.

TurtleBot + OpenMANIPULATOR-X Gazebo Harmonic Differential Drive LiDAR + OAK-D RGB-D Active Development

On this page

Project Statement

The Semantic Fetch Robot is a ROS 2 mobile manipulator operating in a mapped indoor warehouse environment (TurtleBot default Gazebo Harmonic depot world). Given a text command such as “fetch the red bottle”, it autonomously navigates to the relevant region, locates the requested object using open-vocabulary visual detection, grasps it with the mounted OpenMANIPULATOR-X arm, and returns to the operator to deliver the item.

Success state: The robot correctly identifies, grasps, and delivers the requested object in ≥ 75% of trials within a pre-mapped simulation environment, without collisions.

The Problem We Are Solving

Most service robots either move well or manipulate well, but few can do both while also understanding flexible, natural-language requests. Semantic Fetch aims to bridge this gap by enabling a mobile robot to understand what object is being requested, infer where it is likely to be, navigate to it, and bring it back safely.

To achieve this, we combine SLAM-based (Simultaneous Localization and Mapping) navigation with semantically grounded object detection and arm motion planning in a single ROS 2 pipeline. The project focuses on building realistic simulations that serve as a proof of concept before deployment to real hardware.

Project Components

Semantic Mapping

SLAM-built occupancy grid enriched with object detections. Every item detected is registered in a queryable map with its 3D location and semantic label.

Open-Vocabulary Detection

YOLOWorld or CLIP-based detection on the OAK-D camera stream which allows the robot to find objects described in free text without a fixed class list.

Autonomous Navigation

Nav2 stack with SLAM Toolbox handles global path planning and obstacle avoidance. The robot navigates to object locations queried from the semantic map.

Arm Control & Grasping

MoveIt 2 plans collision-free arm trajectories for the OpenMANIPULATOR-X. Grasp poses are computed from RGB-D point cloud data.

Technical Specifications

Parameter Value
Robot Platform TurtleBot Standard + OpenMANIPULATOR-X
Kinematic Model Differential Drive (base) + Serial 4-DOF (arm)
Primary Sensors OAK-D Spatial AI Stereo Camera, RPLIDAR A1 2D LiDAR, IMU
Simulation Engine Gazebo Harmonic (gz-harmonic)
Simulation World TurtleBot default depot world
ROS Version ROS 2 Jazzy Jalisco
OS Ubuntu 24.04 LTS

Current Status

  • Milestone 1 (proposal and architecture) has been completed; details are documented in Milestone 1.
  • Milestone 2 (Implementation, integration, and evaluation in simulation) is completed: details are documented in Milestone 2.
  • Object spawning + full pick-and-place pipeline are planned next.

This site uses Just the Docs, a documentation theme for Jekyll.