Autonomous rally car gets sideways to improve performance in hazardous conditions
Autonomous cars might be able to handle highway conditions pretty well, but they're still no match for humans when things get a little more unpredictable, such as in the dirt. A team at Georgia Tech is looking to change that, using one-fifth sized autonomous rally cars to improve self-driving technology in hazardous conditions.
The scale rally cars are tested on a dirt track, where they're subjected to racing, sliding and jumping at the equivalent of 90 mph (145 km/h). Thanks to clever algorithms, onboard computing hardware and sensors, the cars are able to react to situations and maintain control when a vehicle is near its friction limits.
Rather than using a normal control system, which deals with slippery conditions where the car might be on the edge of its performance envelope the same way as regular highway driving, the Georgia Tech team's car uses a unique approach as it gets close to the limit of grip.
Labeled "model predictive path integral control" (MPPI), the system has been specifically designed to address the non-linear inputs you need to balance a car on the edge. Having integrated large amounts of handling-related information with data on the scale car's dynamics, the system will calculate the most stable way to handle a wide range of scenarios. According to the team, this allows the car to plan and execute the computationally complex handling decisions in real time.
The 3-ft-long (0.91 m), 48-lb (22-kg) cars are custom built, and use a special electric motor providing the right combination of power and weight. They're fitted with two forward facing cameras, an inertia measurement unit and wheel-speed sensors, while power, navigation and computation equipment is held in an aluminum enclosure tough enough to withstand tumbles if a car's confidence overtakes its talent.
In their current guise, the cars are able to test out the team's algorithms without external computation using nothing other than a nearby GPS receiver. That's because the onboard GPU unit can sample up to 2,500 different 2.5 second trajectories in less than 1/60th of a second.
"The algorithms we have developed are able to project into the future what the vehicle is going to do in the next three, or four, or five seconds, and generate approximately two or three thousand possibilities of what is going to happen," Panagiotis Tsiotras, an AE professor who is an expert on the mathematics behind rally-car racing control. "Based on these possibilities, it chooses the best one. This can be done very, very fast. I think we're calculating about 2,000 possibilities every 50 milliseconds."
As well as making the little cars hoon around in a way that would make most petrolheads jealous, this technology has the potential to improve the way autonomous vehicles handle unexpected situations in hazardous conditions.
A video of the truck and interviews with some of the brains behind it is below.
Source: Georgia Tech