Self-driving ATVs — designed at CMU — are made to help in times of disaster

Jimmy Cloutier / Pittsburgh Post-Gazette

As the ATV veered off the gravel road into a patch of knee-high grass, it looked for a moment as if Yifei Liu had lost control of the vehicle.

The quad hit a jagged rock, rattling the computers and sensors that had been packed onboard

But Ms. Lui, a graduate student at Carnegie Mellon University, calmly kept her hands off the steering wheel — confident that the system she helped develop had a handle of the situation. The ATV slowly climbed the grassy hill, then banked right into a small wooded area. It then effortlessly weaved between a pair of trees before rolling to a stop in a clearing.

“Off-road, there’s no clear distinction between a safe trail,” explained Wenshan Wang, a systems scientist at Carnegie Mellon University. 

Researchers from CMU’s AirLab on Thursday held a live demonstration of TartanDriver, their self-driving all-terrain vehicle.

Unlike the far more ubiquitous robo-taxis trained to navigate city streets, the TartanDriver is designed to traverse what researchers call “unstructured” environments where the autonomous vehicle must chart its own path through tricky terrain.

Autonomous vehicles being developed by companies like Pittsburgh-based Aurora and Waymo follow laid-out routes, predictable traffic laws and street markers to travel from point A to point B.

Paths in forests and deserts aren’t as well-defined — if at all.

Consider also the obstacles one might encounter off-road: a steep slope, loose gravel, muddy ditch or patch of tall grass that may or may not be traversable.

“Autonomous driving is a lot more structured than this field,” said Sebastian Scherer, an Associate Research Professor at CMU’s Robotics Institute. “Oh, there’s a car, there’s a person. Avoid them. Here, we’re trying to figure out if we cannot go through this grass. Is this gravel safe? Am I going to get stuck in the mud?”

Using an array of cameras and sensors, the TartanDriver system gathers a trove of data on its surroundings and then judges for itself the safest, most optimal route toward its destination.

To inform those decisions, researchers developed a system that learns not only from observing how humans drive in similar environments but also from its own experience.

“If it goes over some rough area, it says, ‘hey, I shouldn’t drive so fast,’ ” Mr. Scherer said.

The advantage of this approach is that the self-driving ATV can learn to navigate unfamiliar terrain on its own.

Researchers said the TartanDriver had never been to the empty field in eastern Ohio where the demonstration was held.

A monitor behind the driver’s seat displayed a map of what the robot was seeing — and the possible paths it could follow. The red path was the one it had judged to be optimal — one that avoided a precarious rock or muddy puddle.

Mr. Scherer said autonomous vehicles like these could one day ferry supplies to firefighters battling wildfires, assist in search-and-rescue operations or help in clean-up efforts following a natural disaster, when roads are blocked by debris.

Ms. Wang said this technology also will likely control the next rovers sent to the moon and Mars. While operators on Earth must now map routes for the robotic explorers millions of miles away, researchers in the future will be able to provide the next generation of autonomous robots a waypoint and let them figure out how to get there.

At the same time, researchers and developers have to work out several engineering challenges before these autonomous platforms roll into less forgiving environments like coal mines, where heat, dust and tight quarters can throw off sensors or damage electronics.