Overcoming Unexpected Failures using Neurocognitive Multi-abstraction Active Exploration ONR White Paper Tracking Number 23-000005331
Abstract
Approved for Public ReleaseWhen robots are faced with unexpected challenges, they often get stuck; in contrast, humans and animals have demonstrated the ability to recognize when they are stuck and discover new strategies. We will explore the cognitive and psychological mechanisms by which humans and animals learn to solve unexpected challenges, and we will use insights from these studies to develop bio-inspired techniques for enabling autonomous agents to discover new strategies to overcome the unexpected. We aim to develop a unifying framework for problem solving that will both enable autonomous agents to solve problems and also to explain how humans and animals solve problems. This framework will be verified by comparing the framework#s predictions against experimental data from problem solving in humans and animals, and by using the framework to develop methods for autonomous robots to solve problems. Our proposed framework for problem solving in humans, animals, and autonomous agents is structured as follows: Given a new task in an unfamiliar environment, the agent must search for a solution to the task. We believe that the key to planning efficiently is to use the appropriate spatial and temporal abstractions, which are task dependent. Once a plan has been found, the agent must monitor its progress to detect failure (which is inevitable when tackling challenges with unexpected obstacles). After detecting failure, the agent can re-represent the environment using a new spatial or temporal abstraction and replan, or it can perform active exploration to collect more data. We hope to increase our understanding of how humans and animals perform the above processes, and to use this understanding to inspire our development of new methods for autonomous problem solving.In Thrust 1, we will measure the problem solving capabilities of humans and non-human primates (NHPs). We will develop computational models to account for these behaviors and to serve as templates for further algorithmic development. By testing autonomous agents on the same tasks as humans and animals, we can compare the performance of each to determine the limitations of the autonomous agent. These comparisons will guide our development of improved methods for autonomous problem solving.In Thrust 2, we will develop our proposed autonomous problem solving system, which will enable autonomous agents to plan with different spatial and temporal abstractions. If a proposed plan fails to complete the task,a failure management system will choose to either plan with new abstractions or collect new data, which will be used to update our prior models and propose new environment-contextualized abstractions, until the agent discovers a plan that succeeds.To evaluate ourframework, we need to set up a set of tasks in which the agent encounters unexpected challenges. Unfortunately, there do not currently exist any suitable benchmarks for this task. To address this gap, in Thrust 3, we propose todevelop a set of challenges in whichthe agent encounters unexpected obstacles, both in the real world and in simulation, which we will use for both benchmarking (comparing autonomous, human, and animal performance on these tasks) and for evaluation of our method.The team has a strong track record of working with the DoD, including AFOSR, ONR, and DARPA. The capabilities developed in this project will allow autonomous agents to develop new strategies to solve problems that are applicable to the DoD. Strategies for #Retrieve the Target Object# challenge can be used for search and rescue missions or underwater exploration. The #Broken Objects# challenge can enable agents to propose how to repair damaged military equipment. Strategies for the #Unexpected Disasters# challenge can be used to help deal with unexpected challenges on the battlefield. More generally, problem-solving agents that build and plan with causal models could be used to propose new warfighting strategies.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412748
Entities
People
- David Held
Organizations
- Carnegie Mellon University
- Office of Naval Research
- United States Navy