Program aims to develop lifelong learning agents that share knowledge while operating under constraints
Under a new Defense Advanced Research Projects Agency (DARPA) contract for the Shared-Experience Lifelong Learning (ShELL) Artificial Intelligence Exploration (AIE) effort, Aurora aims to develop AI algorithms to achieve life-long learning for agents that learn new tasks in changing environments and share their experiences with each other, while accounting for limitations in communications and hardware configuration.
In the Lifelong Learning (LL) framework, agents continually learn as they encounter new tasks or situations while in the field. The ShELL program extends this approach to many agents that share these continuous experiences among the whole population, thus improving and accelerating the training of each agent in the group. Additionally, the DARPA ShELL agents seek to address size-weight-and-power (SWaP) and computing-constrained platforms with limited communications.
The program will be executed in two phases, with a phase 1 feasibility study and phase 2 proof of concept. Under a subaward of this program, Aurora will collaborate with the Aerospace Controls Lab in MIT’s Department of Aeronautics and Astronautics on a novel AI agent learning methodology and optimize the agents to operate on low SWaP and communications constraints. This is key foundational technology to enable autonomous systems to operate resiliently and robustly in new or unknown environments.
ShELL is the fourth program that Aurora has won under DARPA’s Artificial Intelligence Exploration (AIE) portfolio. Other AIEs are:
- The Intelligent Auto-Generation and Composition of Surrogate Models project, known as Ditto, where Aurora focuses on improving computer-generated design models by integrating artificial intelligence and machine learning to speed up simulations of future military equipment.
- Gamebreaker, where Aurora develops methodologies to assess and affect game balance and game-breaking using a novel adversarial learning approach for efficiently training intelligent agents and learning the game balance models simultaneously.
- Techniques for Machine Vision Disruption (TMVD) where Aurora proved a novel universal attack methodology to disrupt machine vision systems trained to classify militarily relevant objects of interest across multiple image domains.
AIE is part of DARPA’s multi-year investment of more than $2 billion in new and existing programs called the “AI Next” campaign. Aurora is proud to participate in these programs by developing new technologies at the cutting edge of AI and machine learning.