Inspiration
After a disaster, speed saves lives. In the 1980 southern Italy earthquake, 93% of survivors were pulled out within the first 24 hours. Searching also wears rescuers down: at the Surfside condo collapse, crews on 12-hour shifts had to rotate off the rubble every 45 minutes. We wanted robots to take over the repetitive search sweep so rescuers can spend their energy getting people out.
What it does
A group of small, cheap ground robots fan out and search an area at the same time, using patterns that real rescue teams use (grid, sector and radial). When a robot’s camera confirms a person, it stops and radios their estimated position to a laptop base over a LoRa mesh. It needs no GPS, cell service or Wi-Fi.
How we built it
- Hardware: Robot-side laptop (wired to the robot) running YOLOv8 with a USB camera, BNO055 IMU, ESP32 with an L298N motor driver and IR sensor, and a Heltec LoRa radio running Meshtastic.
- Positioning: dead reckoning from IMU heading and calibrated speed × time, combined in a Kalman filter that also tracks how uncertain the estimate is.
- Safety: runs on the ESP32. The motors stop if the robot-side laptop goes quiet for 300 ms, and the robot won’t drive into close obstacles.
- Testing: a simulator with deliberately bad sensors, to check our search patterns before running them on hardware.
Challenges we ran into
Without GPS, small speed and heading errors add up over a long path. The grid pattern drifted up to 5.3 m until we calibrated the speed. We also had to filter out false detections, fit messages into LoRa’s low bandwidth, and merge five people’s separate code into one working robot.
Accomplishments that we’re proud of
In simulation, our sector and radial patterns stayed within about 1.7 m of the true position with no GPS, while covering close to the most area each pattern can. We built a complete loop, from scan command to detection to the base showing the position, on cheap off-the-shelf parts.
What we learned
Search patterns that pass back through the center keep position error low because errors partly cancel out, and calibrating speed did more for accuracy than fancier filtering. We also learned that safety-critical logic belongs on the simplest, most reliable chip, not on the main computer.
What’s next for Group Search Miniature Robots
- GPS for outdoor searches, combined with dead reckoning
- A thermal camera, to find people in low light or when they’re partly hidden
- Multiple robots that split up an area among themselves automatically
- A shared map at the base that shows every find
