Neurala designs and produces software that imitates the actions of the human brain through Deep Learning technology. In 2013, it was awarded a contract with NASA to train the rover, Curiosity, to travel across the Martian terrain. The software Neurala (pronounced neural with an “ah” ending) developed is designed, not just for Mars, but for “planetary exploration.” Graphic Processing Units (GPUs) were installed for visual processing. Curiosity uses machine vision to “lock onto” an object, and remember it. It was designed to “autonomously” move around the Martian landscape, learn the locations of various objects and obstacles, and develop a map of the local terrain. Curiosity was also taught to learn about the objects it came across, and to return to a safe place during serious dust storms.
These activities were all performed with minimal human intervention. Curiosity is not a drone located on Mars, but a “learning, exploring robot.” Rather than driving Curiosity remotely, this clever little explorbot is told where to go, by way of what is seen and mapped. Its cameras send images to NASA, and “locked on” images are mapped, providing a common paradigm and excellent communications. It is told where to go, and then it goes there, no steering wheel required. Curiosity can do this because of Neurala’s brain-inspired, Deep Learning technology. Now, the company plans to adapt this technology to advance the data systems necessary for self-driving cars, remote-controlled drones, and other potentially self-driving vehicles, such as trains, boats, and planes.
Massimiliano “Max” Versace, CEO and co-founder of Neurala, said:
“The Neurala Brain works more like the human brain than any other Deep Learning neural network solution out there. The rigors and demands of building bio-inspired brainpower for space exploration are now benefitting the most exciting innovations here on earth. But, our big differentiation is that our ‘brains’ can think on their own, learn, and act instantaneously. This is essential for applications like self-driving cars, where you can’t rely on network access. And it’s an endearing capability when built into dolls that recognize their young owners by sight and give them a smile.”
Neurala’s Neural Network
Neurala’s software uses a bio-inspired neural network approach to mirror the way a human brain learns to deal with its environment. This “learning software” can be used to support smart products, ranging from delivery drones to toys to self-driving cars. A machine with this software can learn and adapt as it interacts with its environment. Toys can learn to recognize their owner, security cameras can be taught to identify specific threats, and drones can be trained to investigate and diagnose problems at the tops of wind turbines.
Neurala’s programming package is processor-agnostic. It has the ability to operate, without “knowing” a system processor’s design details. Similar to the way NoSQL can process Unstructured Data, processor-agnostic systems can work within a variety of environments. Teams introducing the software into the system will only have to write and install the code once, its flexibility allowing it to work with a variety of different processors within the system.
An additional strength of Neurala is its scalability. According to Roger Matus, Neurala’s VP of Products and Markets:
“An important advantage Neurala has over the competition is its ability to scale to any size, ranging from embedded hardware to servers. We have a system that scales from very small toys, all the way to large servers for inspections and self-driving cars. It sounds very diverse, but all of these have to do with objects learning and observing in the real world, analyzing what they see, and taking action based on what they see.”
Curiosity’s Navigation System
NASA’s rover controllers can tell Curiosity, “Go to this specific location,” and the rover will analyze recorded images, taken by its cameras, and decide on the safest route. The Martian rover sees in stereo, much like human eyes, and then its computer (brain) processes the information, mapping out any geographic hazards or rough terrain. Curiosity will consider all possible paths, and choose the best one, sometimes overriding humans attempting to be helpful.
Neurala’s Deep Learning software also attempts process information in the same way the human eye does. It moves around a high-resolution retina to “quickly” classify objects. The Neurala visual system can efficiently classify, segment, and localize objects. It can also shift into training mode, on demand, learning about new objects in seconds though an advanced data processing system that according to Neurala no one else has at this time. The combined visual system and the navigation system are connected in a way that allows rovers (or other devices) to self-localize without a GPS, and to avoid collisions using an “anticipation process,” rather than a reaction process.
On the 376th Martian day of the mission, Curiosity crossed a previously unexplored depression. The exploration included 10 meters (roughly 10 yards) of “independent navigation” along an obscured stretch of ground. John Wright, a rover controller, said,
“We could see the area before the dip, and we told the rover where to drive on that part. We could see the ground on the other side, where we designated a point for the rover to end the drive, but Curiosity figured out for herself how to drive the uncharted part in between.”
Per NASA’s requirements, Neurala designed and developed its software so rovers could navigate and investigate “other planets” autonomously. To do this, the rovers needed a low-energy processor combined with a software program and data processing system powerful enough to process images, and make intelligent decisions. The processor also needed to be small enough to install in a rover’s body (about the size of a car). Advances in GPUs made this possible.
Innovative Thinking
Neurala created the neural network software used by Curiosity to explore Mars. In designing the software, they had to be fully aware of the limitations on battery life, processing power, and communications. Rather than relying on the traditional mindset of designing a Deep Learning technology for a supercomputer, the Neurala Brain was created to make decisions without support of the Internet. This concept remains intact, as they adjust their software for a wide range of devices on Terra (Earth), including interactive smart toys “that are not” connected to the Internet (eliminating privacy concerns). A Neurala controlled device would not be susceptible to data breaches or hacking, yet still retains the versatility offered by its powerful neural network.
Wheelchair Technology
Neurala also has great potential for people with disabilities. A wheelchair could be told to “Cross the street,” or “Go to the store,” and it would comply by finding the best route for someone traveling in a wheelchair. These wheelchairs would also have the ability to stop, and to move around obstacles, autonomously.
Behind the Scenes
Pelion Ventures, with Motorola Ventures, Sherpa Capital, 360 Capital Partners, Draper Associates, and others have agreed to invest in Neurala for the expansion of their product line and services. Draper Associates and 360 Capital Partners are also providing separate seed funding. The funding will be focused on expanding their customer base and expanding their concept to include other applications.
Ben Lambert, a Senior Associate of Pelion Ventures said:
“Pelion is a firm that believes in a future in which automobiles, robots, drones and other mobile hardware devices will need to navigate the world autonomously and safely, and we believe that Neurala has the right solution to enable that future. Neurala’s executives have a deep technology background and have been leading development efforts in this industry since 2006. As a result, the company is far ahead of its competitors and is at the forefront of the Deep Learning and computer vision space.”
The Future of Space Exploration
While Neurala’s Deep Learning technology has great potential at home, it also has a great deal of potential for space exploration. As we explore our solar system, and the cosmos, this type of autonomy will become more and more useful. The time needed for a problem to be communicated to home, and for instructions to be received by a distant, unmanned ship, makes robotic decision-making an absolute necessity. When robotic submarines begin exploring Europa, or robotic airplanes fly the skies of Mars, autonomous, robotic decision-making will be the only option.
Photo Credit: CNStock/Shutterstock.com