Engineers at Florida Atlantic University are employing an array of advanced technologies to build a robotic dog that can learn from its own mistakes.
By training the machine to teach itself, they’re pushing novelty quadrupeds into the realm of real-world robotic applications.
Dubbed Astro, the four-legged prototype made its debut last week in the Miami suburb of Boca Raton. It possesses petaflops of processing power in an onboard computer brain.
Loaded with deep learning algorithmsOpens a new window , the machine can hone its skills and develop new ones by matching real-time experience to a library of pre-figured scenarios.
The result is a machine that has gone its competitors one better, enabling Astro to travel freely on the battlefield or the shop floor.
Learning on The Fly
Unlike other robot dogs, the machine can train itself to carry out the military and public safety tasks its creators at the university’s Machine Perception and Cognitive Robotics Laboratory envision it performing.
Indeed, the lab’s interdisciplinary team believes it has created a categoryOpens a new window of a machine that can learn on the fly and adapt its behavior to local conditions.
Until Astro arrived, major advances pioneered by longtime robotics maker Boston Dynamics and others centered on leg-borne machines that could correct for uneven terrain and regain their footing should they topple over.
Named for a Cartoon Canine
Researchers in disciplines ranging from psychology to mathematics had a hand in creating AstroOpens a new window , which bears the name of a cartoon canine from the 1960s cartoon television series “The Jetsons.†A pop-culture reflection of Space Age research and development of the ‘60s, the show chronicled the lives of a futuristic family, including its collection of robotic helpers.
In keeping with the theme, Astro’s computer brain is equipped with Jetson-family processors from Nvidia, a California-based chip designer. They form the guts of a graphics engine capable of four trillion calculations per second.
The platform, housed in a 3D-printed head modeled on a Doberman Pinscher, processes input from an array of sensors that register light and sound and measure altitude and temperature.
The robot’s cameras permit radar imaging, while a directional microphone helps it to cogitate spoken words into executable demands.
Of Guns and Bombs
The 100-pound unit is outfitted for military applications, including detection of guns and the residue from gunfire, as well as bombs and improvised explosive devices.
Paired with first responders, the robot can be used in search-and-rescue missions, similar to the ones that occur when hurricanes in Florida and elsewhere destroy homes and displace residents.
Astro also can work as a service dog for the blind. And it can serve as a medical monitor, collecting and transmitting patient data from wearable sensors.
The robot can search databases to recognize faces as well as detect and relate information and images about environmental conditions in remote areas. And, like real dogs, Astro can pick up sounds at frequencies that exceed the limits of human hearing.
Better Than a Rolled-up Newspaper
The team from the university lab that built Astro liken the robot to a puppy in training.
Astro responds to verbal commands but the researchers hope that the computer-enabled capabilities will help it to react to hand signals and execute commands uttered in languages other than English.
While production is still in the future, the researchers believe a market will become available for more advanced quadruped robots. Massachusetts-based Boston Dynamics and Ghost Robotics of Philadelphia, Pennsylvania, as well as China’s Unitree and ANYbotics of Switzerland all offer quadrupedsOpens a new window for military and industrial applications.
But none possess Astro’s capacity for self-correction when it comes to performing tasks.
To get there, those companies and others may well revert to real-world research. A team of data scientists from the University of Washington last year produced a paperOpens a new window that demonstrates the value of dog behavior for modeling machine-learning algorithms.