Researchers from Carnegie Mellon University’s Robotics Institute and spin-off company Sensible Machines recently flew an autonomous drone through dark, smoke-filled compartments to map fires and locate victims. The demonstration was part of an Office of Naval Research (ONR) project called Damage Control Technologies for the 21st Century (DC-21).
“With the micro-flyer, we wanted to show that it could autonomously navigate through the narrow hallways and doors – even in dense fire smoke – and locate fires,” said Thomas McKenna, ONR’s DC-21 program manager. “It succeeded at all those tasks.”
In a real emergency, information gathered by the micro-flyer would be relayed to a large humanoid robot, the Shipboard Autonomous Firefighting Robot (SAFFiR). The robots would work with human firefighters to put out fires and evacuate casualties.
“Flying autonomously through narrow doorways in darkness and smoke poses a number of technical challenges for these small drones,” said Sebastian Scherer, systems scientist at CMU’s Robotics Institute. “But this capability, known as ‘fast lightweight autonomy,’ will have numerous applications beyond shipboard fires, such as investigation of building fires and inspection of hazardous chemical tanks and power plant cooling towers.”
The challenges begin with the size of the drones. To fit through the 26-inch-wide hatches of the ex-USS Shadwell, a ship in Mobile, Ala., used to test firefighting techniques, Sensible Machines built a quadrotor just 23-inches wide and 12-inches high – a bit smaller than those typically used by hobbyists.
The drone was able to negotiate the tight spaces. But its smaller rotors reduced its efficiency, limiting its flight time to about five minutes. Sensible Machines is now building a drone that’s just 16 inches wide but replaces the four rotors with a single ducted fan with two larger, counter-rotating propellers. Scherer said the larger rotors work more efficiently and are anticipated to boost flight time to 30 minutes.
The primary sensor used by the drone to build its map of fire areas is a RGB-D camera, or depth camera, similar to that of a Kinect game controller. “It actually works better in the dark,” Scherer noted, because there’s less ambient light to interfere with the infrared light the camera projects. Other researchers have tried using depth cameras to do mapping, he said, but have had limited success because they still rely primarily on visual features, with depth information used supplementally.
“We flipped it around, using mainly the depth camera to build our maps,” Scherer said. In addition to the RGB-D camera, the drone uses a forward-looking infrared (FLIR) camera to detect fires and people and a downward-facing optical flow camera to monitor the motion of the drone itself.