Teaching autonomous vehicles to learn by doing
A new project to teach ground robots to learn by doing, rather than responding to verbal commands, will improve how autonomous systems move through rugged and unfamiliar terrain.
Researchers from the Army Research Laboratory and the University of Texas at Austin are working on software that can learn from human demonstrations and continually update estimates of its ability to perform the new behaviors and meet system specifications, ARL officials said in a statement. The software is the Lab’s Autonomy Software Stack -- a suite of algorithms, libraries and software components that intelligent systems use for navigation, planning, perception, control and reasoning when performing specific tasks.
The plan is to develop an autonomous system that interactively learn from soldiers, who can then, based on their confidence in the robot’s ability to apply learned behavior to new tasks, make better choices about how to use the autonomous systems. This capability could open opportunities for robotic systems that learn and execute new behaviors in operational environments without coming home again to be checked, said Army autonomous systems researcher Craig Lennon.
"Allowing autonomous systems to learn new behaviors after fielding would be a significant step forward in how the Army trains, approves and integrates autonomy into units" he said.
"Suppose the robot has already learned procedures for crossing a danger area but has never crossed a river with the load that it's now required to carry," he said. "During mission rehearsal, the soldier finds a river in a friendly area, demonstrates the river crossing by tele-operating the fully-loaded robot across and repeats demonstrations until the system can provide assurance it has learned how to safely execute the behavior while conforming to the tactical procedures for crossing a danger area. Now the soldier can take it on the mission and it can get itself across the river."
Soldiers could also use demonstrations to teach robots how to operate when they’ve been damaged, Lennon said.
In the first year of the project, researchers will test the new software on simulated systems. By mid-2022, they expected to transfer tests to a Clearpath Warthog -- a squat vehicle, roughly five feed square, that can carry a payload of up to 600 pounds.
In May, ARL announced it had awarded research partners on a similar project $2.9 million for first-year funding to develop robotic and autonomous aerial and ground systems that can enter into an unfamiliar, unmapped, communications-denied areas, make sense of the environment and perform safely and effectively.
Connect with the GCN staff on Twitter @GCNtech.