Army research tackles tough terrain for Soldier-robot teams, builds legacy for future innovation

September 09, 2019

By U.S. Army CCDC Army Research Laboratory Public Affairs

ADELPHI, Md. (Sept. 9, 2019) -- In today's warfighting environment, Soldiers need trustworthy teammates more than ever to ensure superiority on the battlefield. Army researchers are working to provide this necessity in the form of reliable robot teammates that will have the ability to traverse through the toughest of terrains.

Imagine Soldiers on a mission in an unstructured environment such as an area with dense vegetation and nothing but dirt roads that have been muddied from a recent rainfall and covered in debris.

Having no prior knowledge of the state of its surroundings, how would a robot quickly perceive and classify the environment, reason about the different quality of surface and potential hazards in its way, and then adapt its maneuver behavior to adjust to these changing conditions?

These are some of many challenging questions that researchers at the U.S. Army Combat Capabilities Development Command's Army Research Laboratory and industry, academic and government partners have been pursuing for almost 10 years under the Robotics Collaborative Technology Alliance, or RCTA.

The initiative is coming to a close in October with a capstone event. The RCTA is a collection of individual research projects that will come together to show the world what the future of Army robotics looks like.

Current RCTA alliance members include CCDC ARL, General Dynamics Land Systems, Carnegie Mellon University, Massachusetts Institute of Technology, Florida State University, University of Central Florida, University of Pennsylvania, QinetiQ North America and the Caltech/Jet Propulsion Lab.

In June, alliance members convened at Camp Lejeune, North Carolina, for a capstone experiment.

According to ARL computer scientist Dr. Maggie Wigness, who has been a part of the alliance for the past four years, the June RCTA capstone experiment took place in an environment designed to represent a village in the developing world, with many small building structures, gravel and dirt roads, and brought together a number of research successes into an integrated system that allows a mobile robot to operate autonomously and reliably in this unstructured environment.

"The different research components, such as scene segmentation, object detection, or route planning, that make up the integrated system have been tested and evaluated individually using benchmark datasets available to the research community, in simulated environments, or within very simple environments, e.g. large open areas without obstacles," Wigness said. "Only the integration of these components provides the necessary intelligence to maneuver, navigate and accomplish dynamic tasks in unstructured environments."

Researchers said they will take what they learned from the June experiment and adjust where necessary to be ready for the capstone demonstration in October.

Why the focus on unstructured environments, you ask? And why not just buy a commercially available robot to complete tasks for Soldiers?

"Much of the commercial research in autonomous navigation and maneuver is focused on highly structured environments, e.g. urban cities with street signs and lane markings," Wigness said. "It is extremely difficult to use these same approaches that rely on structural cues, in an environment that provides little to none of this context. Since our goal is to advance technology that supports our warfighters, it is essential that the RCTA focused the research effort on unstructured environments because this is where our warfighters operate."

One of the research projects under the RCTA that has set out to enable robots to effectively maneuver in unstructured environments involves legged locomotion.

RCTA researchers have developed a dynamic quadruped vehicle that is all-electric, human scale, quiet, autonomous, moves at roughly Soldier walking speed and can carry a meaningful payload in addition to its sensors, computers and battery. The vehicle, a robotic research platform, is called LLAMA, for "Legged Locomotion and Movement Adaptation."

According to Dr. Larry Matthies, senior research scientist at the Jet Propulsion Laboratory and a principal investigator for the RCTA, one generation of the LLAMA is complete, which has a top speed of 0.8 meters per second and can carry a 5-kilogram payload.

A second generation is currently under development and is expected to achieve 1.0 meters per second and carry a 10-kilogram payload. The researchers expect to demonstrate this second-generation LLAMA at the October event, which will be able to perform autonomous navigation around obstacles to destinations a moderate distance away, using onboard depth and dead-reckoning sensors.

Matthies noted that designs exist for a third generation that leverages more advanced manufacturing techniques.

"A core innovation of LLAMA is the use of custom, large diameter, low gear-ratio actuators that have no mechanical springs," Matthies said. "This design can achieve variable compliance in software, which enables the use of sophisticated control techniques for adapting leg stiffness to unknown terrain with variable hardness, which is encountered in real-world scenarios that vary from mud, to gravel to sand/rock mixtures, to solid, paved surfaces. In contrast, most commercial quadrupeds are mechanically tuned, via springs, to a narrower set of operating conditions."

According to the researchers, wheeled and tracked vehicles are limited in the variety of terrain they can negotiate, and cannot keep up with walking Soldiers in all terrain.

"Legged locomotion systems are intended to overcome this limitation," Matthies said. "Previous legged locomotion systems that were not omnidirectional had limited agility in confined spaces, such as indoors or in cluttered outdoor conditions. LLAMA reduces these limitations while being able to carry a payload that may be reconnaissance sensors, supplies for Soldiers, or other equipment such as a manipulation system."

Wigness noted that in addition to being able to maneuver in unstructured environments just as a warfighter, a reliable robot teammate will need to be able to adapt and learn when there are environment or mission changes such as roadblocks, debris and people in the area.

By integrating several traversal behaviors that were developed through collaboration with Carnegie Mellon University, robots will have the ability to context switch while navigating. A Clearpath Husky, a commercially available wheeled research platform, with this technology will be showcased at the capstone event.

"A reliable robot teammate can't have a single traversal mode," Wigness said. "It needs to adjust its behavior based on the context of the environment. When pedestrians are perceived, the robot operates using socially compliant navigation technology, developed collaboratively with CMU, but when the robot believes there is danger in the area like a weapon, it can switch to navigate in a more covert behavior."

According to Wigness, technologies developed in the RCTA, such as maximum entropy inverse reinforcement learning, have also led to on-line behavior learning through human intervention. This allows a warfighter teammate to demonstrate novel behaviors during operation, which the robot uses to adjust its maneuver behavior in the field with minimal interruption to its task execution.

"This ability to adapt quickly is essential in order to maintain operational speed within a warfighter unit," Wigness said.

The robots the researchers will use during the capstone demonstration will have the ability to visually perceive the environment, e.g. detect objects and differentiate terrain types, and will use this perception information to make navigation and behavior decisions, such as avoiding rough and potentially dangerous terrain that may be difficult to traverse.

One of the RCTA research projects that focuses on perception to enhance navigation and thus maneuvering deals with terrain perception for maneuver in unstructured environments.

"In the early years of this RCTA program, algorithms for passive 3-D perception and robot visual dead-reckoning with onboard stereo cameras (stereo vision) developed by JPL and funded by ARL were adapted to exploit compact, low-power, embedded computer systems to enable perception on small robot platforms," Matthies said. "At the time, this was the most power-efficient stereo vision implementation in existence. We also extended this to an extremely efficient algorithm for computing optical flow, and demonstrated the ability to detect and track moving objects with the combination of depth perception and optical flow."

According to the researchers, these capabilities have now become commercially valuable, which inspired a larger research community to work in this area, leading to the development of dedicated hardware implementations of depth cameras and visual deadreckoning systems as commercial products, using further evolved algorithms.

Another research front focuses on the challenge of enabling robots to infer how well they are likely to be able to push through tall vegetation ahead of the vehicle.

"We use a local, 3-D, voxel-type map around the robot to record information from a multi-return LiDAR about local vegetation density, as well as vegetation resistance as measured by a force-sensing bumper on the front of the robot and motor current sensors to measure tractive force," Matthies said. "With these measurements, the robot can estimate the force exerted by terrain that resists motion. These estimates and the vegetation density information from the voxel model are input to a deep learning model that learns to predict the resistance that will be experienced, based on perception of the vegetation ahead of the vehicle instead of on physical contact."

Matthies stated that robots operating in unstructured environments will encounter a wide variety of terrain with a wide variety of trafficability characteristics.

"It is impossible to manually preprogram robot perception systems with the ability to understand the trafficability of the diversity of terrain they will encounter; robots must instead be able to learn about terrain trafficability from their own experience and to adapt their behavior accordingly," Matthies said. "In particular, mobility in tall vegetation is a challenge for human-scale and smaller robots, because vegetation is more of a barrier to mobility the smaller the vehicle."

The researchers said this is a difficult problem that will only be partly solved by the time of the capstone.

Another perception research project that has progressed through the RCTA deals with human perception for maneuver around people.

Advancements in human perception and leader-follower include the ability to localize, track and navigate around humans in complex terrain.

"Human perception uses passive stereo vision, and is highly robust to human pose, significant occlusion from terrain or other pedestrians and egomotion of the vehicle," Matthies said. "Our current capability can track multiple contacts while maneuvering through heavy brush and vegetation, enabling human-robot maneuvers in realistic off-road terrain. The model is based on monocular instance segmentation using deep learning from stereo and filter-based multiple instance tracking."

Through the RCTA, the greatest minds in robotics research have joined forces to progress towards a reality for future warfighters where they are paired with a reliable teammate that can keep up with the demands of their missions, thus unburdening and further protecting them in combat.

"This research in maneuvering in unstructured environments is so vital to the future Army and Soldier and will thus continue after the RCTA under ARL's Artificial Intelligence for Maneuver and Mobility Essential Research Program," Wigness said. "Our nation's land power dominance will continue to rely on significant science and technology advances that ensure competitive advantage for maneuver forces, which ARL and its partners will relentlessly continue to pursue now and for years to come."




The CCDC Army Research Laboratory (ARL) is an element of the U.S. Army Combat Capabilities Development Command. As the Army's corporate research laboratory, ARL discovers, innovates and transitions science and technology to ensure dominant strategic land power. Through collaboration across the command's core technical competencies, CCDC leads in the discovery, development and delivery of the technology-based capabilities required to make Soldiers more lethal to win our Nation's wars and come home safely. CCDC is a major subordinate command of the U.S. Army Futures Command.

 

Last Update / Reviewed: September 9, 2019