DARPA’s Defense Sciences Office is soliciting ideas to explore the question of how, and to what extent, can we assure foundation model-enabled robots will behave only as directed and intended?
DARPA is seeking investigations into innovative methods and approaches that will lead to assurances about the behavior of robots that use functional methods (FMs), with particular emphasis on robots (of any type) that receive commands in natural language; operate in unstructured, open-world environments; and/or incorporate the FM in closed-loop decision making. Methods and approaches may consider all phases and components of the design and training of the FM and the FM-enabled robot as a system. Approaches of particular interest yield assurances (which may be probabilistic) with clear assumptions and are likely to extend as far as possible to complex robot instructions and behaviors in open environments. Methods and approaches that provide assurances subject to minimal additional supervision from a human operator in real time are also of particular interest. Negative results are also in scope, but negative results should be justified as applicable to current or future FM-enabled robots with state-of-the-art performance and capabilities.
USD$100,000 – $300,000 is available for projects of up to 12 months.
Abstracts are due 13 January 2025 and will be assessed on a rolling basis, until 13 January or available funding is exhausted, whichever occurs first.
An information webinar will be held 22 October 2024. Register here by 21 October. Non-US citizens must submit DARPA Form 60 (Visit Request) by 16 October.
This Advanced Research Concepts Opportunity is issued under DARPA Exploration Announcement DARPA-EA-24-01.
Documents
ARC Opportunity – SAFRON: DARPA-EA-24-01-05
Exploration Announcement – ARC: DARPA-EA-24-01-Amendment-01
