Follow: Person Following Quadrupedal Robot
Robotics · Computer Vision · Autonomy

Follow: Person Following Quadrupedal Robot

← All Projects

Overview

Follow is an assistive quadrupedal robot designed to identify, lock on to, and follow a person autonomously. Built with a philosophy of smart engineering using minimal hardware and simple algorithms, Follow demonstrates how capable robotic systems can be achieved without complex sensor arrays or expensive components. The robot serves as a service platform for events and conferences — guiding attendees, showcasing robotics technology, and laying the groundwork for future conversational interaction.

Hardware

Robot Platform

The project uses the Unitree Go1 quadrupedal robot as its mobility platform. The Go1 provides a stable, agile base capable of walking across various terrains and indoor environments.

Camera Setup

A custom-mounted RGB camera is attached to the front of the Go1 using screws. This single camera serves as the robot's only perception sensor, reinforcing the project's minimal-hardware approach.

Compute

All onboard processing is handled by a Jetson Nano NX, providing sufficient GPU-accelerated compute for real-time person detection and re-identification.

Software Architecture

Vision Pipeline

1. The RGB camera captures frames in real time.

2. YOLO detects all people in the frame, producing bounding boxes.

3. FastReID extracts appearance feature embeddings from each detected person and compares them against the locked-on target to maintain persistent identity tracking.

Motion Control

The robot follows the person by sending velocity commands based on the person's position within the camera frame. A PID controller ensures smooth and stable tracking behaviour. Distance to the person is estimated using the width of the detection bounding box as a proxy. The robot maintains a safe distance of approximately 1 meter.

State Machine

Standby

The default state. The robot waits passively for a user to initiate interaction.

Follow

Activated when a user verbally says "Follow" while raising their hand. The robot locks on to the person using YOLO detection and FastReID feature embeddings. It then tracks and follows the person, maintaining approximately 1 meter of distance.

Lost

Triggered when the tracked person is no longer visible. The robot enters a foraging search behaviour: performs a 360-degree rotation with deliberate stops, analyses each direction for the highest probability match, and moves toward the most likely direction.

Demonstration

Follow was demonstrated live at Zuitzerland, a pop-up village event hosted in Laax, Switzerland, in June 2025.

Challenges

  • Overexposure from lighting — Bright or uneven lighting can degrade detection performance.
  • Obstacle occlusion — Objects or people passing between the robot and target cause temporary loss.
  • No spatial mapping — The robot doesn't currently build an occupancy grid, limiting search efficiency.

Future Work

  • Spatial mapping and occupancy grids — Environment mapping for smarter search navigation.
  • Obstacle avoidance — Detecting and navigating around obstacles.
  • Conversational interface — Connecting a conversation module to verbally guide users.
  • Improved lighting robustness — Handling overexposure and challenging conditions.
Tech Stack
Unitree Go1Jetson Nano NXYOLOFastReIDROSPID Controller
Team
Ronald Lanton
Concept, Robotics, Control System, Ideation, Vision Model, State Machine, Network Architecture
Timeline
Nov 2024 — Started
Jun 2025 — Live demo at Zuitzerland
Ongoing — Continuous improvements