Sithija Ranaraja

All projectsMobile robotics, HRI

Multi-floor receptionist robot

A semi-humanoid that guides visitors across a whole building

The receptionist robot, from design to trials in real buildings.Hardware

Most indoor robots live on a single 2D map. This project extended 2D SLAM and Nav2 into a multi-floor framework so a receptionist robot can take visitors anywhere in a building. The base is custom sheet metal carrying an RPLIDAR A2, an Intel RealSense D435 and a Jetson AGX Xavier.

Type
Hardware
Period
2024 – 2025, continued as Receptionist 2.0
Role
Team lead: robot design, mapping, navigation and the multi-floor framework
Stack
ROS 2 Humble, Nav2, SLAM Toolbox, BehaviorTree.CPP, MoveIt, Gazebo, Jetson AGX Xavier, RPLIDAR A2, Intel RealSense D435, ReSpeaker, C++, Python
Links
GitHub repository Multi-floor navigation video Lift simulation video Paper: occlusion-aware obstacle prediction (arXiv)

Challenge

Lifts break 2D navigation. The robot has to call the lift, wait, enter, ride, work out which floor it is on and switch maps, with people in the way.

Approach

Behaviour trees (BehaviorTree.CPP) sequence the whole lift routine. I developed it in Gazebo first, with lift plugins written for the project. A ReSpeaker array handles voice and direction of arrival. A human-aware navigation layer with a replicated goal-prediction method keeps the robot out of people's paths. MoveIt plans the arm.

Result

A prototype tested in real buildings. The software is split into ROS 2 packages for navigation, perception, speech, behaviour and the HMI.

What I built

  • Multi-floor navigation framework built on 2D SLAM and Nav2
  • Custom Gazebo plugins for elevator simulation
  • Behaviour-tree automation of the complete elevator sequence
  • Human goal prediction to improve crowd navigation
  • Voice interaction through a ReSpeaker microphone array
  • Design calls: local model over cloud LLM, QoS tuning on critical sensors, behaviour trees over state machines, LiDAR SLAM over visual SLAM