Cambridge-based Charles River Analytics (CRA) has been funded by the US Defense Advanced Research Projects Agency (DARPA) under the OFFensive Swarm-Enabled Tactics (OFFSET) programme to develop Meta-Reinforcement Learning Innovation for Robust Swarm Tactics (MERLIN-RST). MERLIN applies a meta-reinforcement learning approach to discover and learn novel swarm tactics.

Through biology-inspired algorithms and deep machine learning, tactics in the OFFSET programme help swarms achieve and adapt to mission objectives. However, human programmers cannot support the evolution of these tactics quickly or accurately enough - numerous variants in adversary strategy and evolving environments make manual development a challenge that may lead to system failures or ‘brittle tactics.’ MERLIN therefore uses a meta-reinforcement learning approach so swarms can autonomously learn new tactics.

MERLIN explores a rich, complex and unpredictable search space that describes the many possible forms a tactic can take and automatically optimise it to a wide range of mission goals,” commented Michael Harradon, Senior Scientist at CRA and Principal Investigator on the MERLIN effort. “We’re laying the groundwork for a highly flexible design process that can dramatically improve swarm effectiveness.”

Under the MERLIN effort, the company is augmenting OFFSET simulators to model swarm performance in a wide range of challenging urban environments as part of the OFFSET’s Fourth Swarm Sprint, building on recent successes under the programme. The Swarm Algorithms and Tactics for Urban Reconnaissance and Isolation (SATURN), CRA developed capabilities that give heterogeneous swarms of unlimited size resilient behaviour while achieving mission objectives. The Ecological User Interfaces for Rapid Operational Teaming with Autonomous Swarms (EUROPA) development helps operators better control swarms in urban operations by providing novel, multimodal user interfaces tailored to tactical operational environments.

CRA’s OFFSET participation complements other techniques and applications the enterprise has developed to support autonomous and unmanned systems, such as the Multi-modal Interface for Natural Operator Teaming with Autonomous Robots (MINOTAUR) interface for controlling robotic leader-follower systems, the Advanced Mission Planning Tools (AMPT) framework for supervising unmanned vehicles and the Compliance in Real-Time Onboard Watercraft for Safe Navigation using Environs-Modeling, Sensor-Fusion and Tracking (CROWSNEST) maritime traffic awareness solution for an unmanned surface vessel.

MERLIN uses a meta-reinforcement learning approach so swarms can autonomously learn new tactics. (Image: CRA)

- Mönch Publications - Latest Issues -

We use cookies to improve our website and your experience when using it. Cookies used for the essential operation of the site have already been set. To find out more about the cookies we use and how to delete them, see our privacy policy.

I agreeMore Information