Bio-Inspired Algorithms Walk AI Robots Through Learning Process

Machine learning AI-min

When it comes to getting artificial intelligence systems to be more “human,” robots apparently have to learn to walk before they can run.

The Pentagon’s top research arm recently hit a milestone in enabling a machine to learn without having to go through a lot of reprogramming at each stage, like they do now. And they did it with a robotic limb teaching itself to walk.

The limb, with animal-like tendons and controlled by a bio-inspired AI algorithm, was taking steps after a mere five minutes of “unstructured play,” during which it learned about its own construction and the surrounding environment, according to a release from the Defense Advanced Research Projects Agency (DARPA). Those small steps, which included righting itself after being knocked off balance, amount to something of a leap in AI development according to researchers working under DARPA’s Lifelong Learning Machines (L2M) program.

“We’re at a major moment of transition in the field of AI,” said Dr. Hava Siegelmann, program manager in DARPA’s Information Innovation Office (I2O). “Current fixed methods underlying today’s smart systems will quickly give way to systems capable of learning in the field.”

Working with a team led by Francisco J. Valero-Cuevas of the University of Southern California’s Viterbi School of Engineering, the work furthers L2M’s overall objective of getting machines to learn and adapt on the fly. “These abilities are necessary, for instance, for complex systems like self-driving cars to become truly functional,” Siegelmann said.

It also would apply to any number of military systems, from swarming drones and ground vehicles, to AI-powered cybersecurity, which are among the goals of DARPA’s $2 billion AI Next program.

Despite recent advances in machine learning, systems are still almost entirely tethered to their training. After detailed, repeated sessions in a number of established scenarios, an AI system will know what to do. But it will be confounded by the unexpected. In cases like that, DARPA points out a machine would have to be taken offline, reprogrammed, and retrained, which not only takes a lot of time and effort, but doesn’t do any good in a real-life, real-time situation.

L2M, which kicked off in 2017 and covers about 30 research groups working via grants and contracts of different duration and sizes, is intended to change that. “The L2M program’s prime objective is to develop systems that can learn continuously during execution and become increasingly expert while performing tasks … without forgetting previous learning,” Siegelmann said. “Though complex, it is an area where we are making significant progress.”

In a paper published at Nature Machine Intelligence, Valero-Cuevas and his team of USC doctoral students explain how they created a three-tendon, two-joint limb like those of invertebrates and a “biologically plausible algorithm” called G2P (general to particular) to enable what they refer to as “few-shot” learning. Like invertebrates, the robotic limb learns from just a few trail-and-error attempts, rather than from a heavily modeled program. Further work along these lines, they write, “could imbue robots with the enviable versatility, adaptability, resilience, and speed of vertebrates during everyday tasks.”

The military has plenty of uses for machines that can truly learn on the job, from aircraft and reconnaissance systems to those that can conduct electronic warfare and identify unknown radio frequency signals. But such improvements in machine learning also will have an impact on everyday applications, such as self-driving cars that actually are safer than the one driven by that guy just ahead who doesn’t use his turn signal, or language translation, image identification, or even customer interactions and housekeeping.

The L2M project, which combines hardware and software innovations, could help get that goal closer to reality.