Army researchers use human teaching to improve navigation in autonomous systems

In the future, it may only require a soldier and a game controller to teach robots how to surpass humans.

At the Army Research Laboratory of the U.S. Army Combat Capability Development Command and the University of Texas at Austin, researchers designed an algorithm that allows autonomous ground vehicles to improve their existing navigation systems by observing human driving. The team tested its method on the Army’s experimental vehicle Clearpath Jackal, which uses a demonstration to perform Adaptive Planner Parameter Learning From Demonstration (APPLD).

Army researcher Dr. Garrett Warnell said: "Using the APPLD method, active soldiers in existing training facilities will be able to improve the autonomous navigation system by simply operating the vehicle normally." "Technology like this will help. The Army’s plan to design and deploy next-generation combat vehicles capable of autonomous navigation in off-road environments has made an important contribution."

The researchers combined the demonstration algorithm with machine learning in more classic autonomous navigation systems. APPLD does not completely replace the classic system, but learns how to adjust the existing system to behave more like a human demonstration. Warnell said that this paradigm allows the deployed system to retain all the advantages of traditional navigation systems, such as optimality, interpretability, and safety, while also making the system flexible and adaptable to new environments.

Warnell said: "Using a human driving demonstration provided by the daily Xbox wireless controller, APPLD can learn how to adjust the vehicle's existing automatic navigation system in different ways according to the specific local environment." "For example, in a narrow corridor, The driver slows down and drives carefully. After observing this behavior, the autopilot system learns to reduce its maximum speed and increase its computational budget in similar environments. This ultimately allows the vehicle to successfully navigate other narrow corridors that have previously failed Autonomous navigation in China."

Click the link to watch the video ：https://mp.weixin.qq.com/s/pvp443sIkYzkvZliRQxNPw

Dr. Peter Stone, chairman and professor of the Austin Robotics Alliance, said: “APPLD is another example of the unique collaboration between Austin and the Army Research Laboratory that has contributed to the continuous growth of research results.” “By placing Dr. Warnell full-time in At Austin, we can quickly discover and solve cutting-edge scientific research problems."

The team’s experiments show that after training, the APPLD system can navigate the test environment faster and reduce failures compared to traditional systems. In addition, a trained APPLD system is generally faster than a trained person to navigate the environment. The peer-reviewed journal IEEE Robotics and Automation Letters published the team’s work: APPLD: Learn adaptive planner parameters from a demonstration.

"From a machine learning perspective, APPLD is in sharp contrast to the so-called end-to-end learning system that tries to learn the entire navigation system from scratch." "These methods often require a lot of data and can lead to behavior that is neither safe nor secure. Robust. APPLD uses a well-designed control system part, while focusing its machine learning results on the parameter adjustment process, which is usually done based on a person’s intuition."

APPLD represents a new paradigm in which people without expertise in robotics can help train and improve the navigation of autonomous vehicles in various environments. Rather than a small team of engineers trying to manually adjust the navigation system in a few test environments, in fact, an unlimited number of users will be able to provide the system with the data needed to adjust themselves to an unrestricted environment.

Dr. Jonathan Fink, an Army researcher, said: “Usually, for each new deployment environment, the current autonomous navigation system must be manually retuned.” “This process is very difficult and must be conducted by personnel trained in robotics. It requires trial and error until the correct system setting is found. On the contrary, APPLD automatically adjusts the system system by observing manual driving. Anyone with experience in video game controllers can do it. During the deployment process, APPLD also Allow the system to readjust itself in real time as the environment changes."

The Army’s focus on modernizing next-generation combat vehicles includes designing optional manned combat vehicles and robotic combat vehicles that can navigate autonomously in an off-road deployment environment. Although soldiers can navigate in these environments with current tanks, the environment is still full of challenges for advanced autonomous navigation systems. APPLD and similar methods provide new potential ways for the Army to improve its existing autonomous navigation capabilities.

In addition to being directly related to the Army, APPLD also provides an opportunity to bridge the gap between traditional engineering methods and emerging machine learning technologies, thereby creating powerful, adaptive and versatile mobile robots in the real world.

In order to continue this research, the research team will test the APPLD system in various outdoor environments, employ soldier drivers, and try a variety of existing automatic navigation methods. In addition, the researchers will investigate whether including other sensor information (such as camera images) can lead to learning more complex behaviors, such as adjusting the navigation system to operate under various conditions (such as different terrain or the presence of other objects).

References: X. Xiao, B. Liu, G. Warnell, J. Fink and P. Stone, "APPLD: Adaptive Planner Parameter Learning From Demonstration," in IEEE Robotics and Automation Letters, vol. 5, no. 3, pp . 4541-4547, July 2020, doi: 10.1109/LRA.2020.3002217.

The Clearpath Jackal unmanned vehicle is a small, fast entry-level field robot research platform. It has an onboard computer, GPS and IMU fully integrated with ROS, with autonomous functions out of the box. The maximum speed of Jackal can reach 2m/s, and the linear acceleration can reach 20m/s2. With a load of 20kg, the maximum climbing angle can reach 40°. The company has a small amount of stock, friends in need are welcome to contact us.