REPRISM Flexible Embodied Problem-Solving by Manipulating the Representational Prism
Abstract
Existing robot systems are remarkably brittle, and are consequently limited to highly constrainedlaboratory settings; they are not suitable for, and would immediately fail in, realworlddeployments. By contrast, humans and some animals display remarkable flexibilityin problem-solving#in problem-solving strategy, creative use of available resources,task-level exploration, and in the recruitment, adaptation, and occasional abandonment ofexisting knowledge. Only when robots are able to show the same level of flexibility and resourcefulnesswill they be sufficiently robust and autonomous for real-world deployments.Our core hypothesis is that truly flexible problem-solving can only be achieved by a robotthat devises, adapts, and exploits its own structured representation of novel tasks#whichwe call a representational prism#that enables efficient problem solving. The representationalprism contains four elements. The robot constructs an abstract decision processthat models the task by combining a perceptual lens that selects relevant perceptual inputswith a motor lens that selects complementary motor skills. The resulting abstract decisionprocess is operated upon by a problem-solving strategy to produce a control policy byexploiting structure present within it. Overseeing problem-solving is a metacognitive processthat constructs the prism and monitors overall performance, modifying the perceptualor motor lenses and solution strategy as necessary.Such a robot will be able to adaptively include only relevant sensorimotor resources to minimizecomplexity and discover task structure, selectively import task-general backgroundknowledge, and rapidly specializing to task-specific experience. It will be able to selectfrom a portfolio of problem-solving strategies, specialize existing solvers, or discover newones. It will be able to speculatively impose and exploit structural assumptions to supportefficient solution methods. It will be able to estimate task learning time and monitorprogress, to determine when to change strategies, consider the task solved, or give up. Finally,the robot will beable to change its representation at any time, expanding or reducingits scope as data indicates, or selectively recruiting sensorimotor resources and modifyingstructural assumptions as a means of task-level exploration. Such a robot would achieve aqualitatively new level of problem-solving flexibility.Our interdisciplinary team, consisting of four brain scientists and four roboticists, proposesa sustained, coordinated, and multidisciplinary effort that realizes every component of theprism based on insights drawn from problem-solving and play via cognitive science, neuroscience,and animal behavior, and in light of its role in a complete integrated robot system.Our goal is to realize robots that can be deployed to robustly problem-solve in complex,challenging, unexpected, and possibly adversarial situations. The proposed work is FundamentalResearch, and is to develop technology for both military and civil application.Approved for public release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412603
Entities
People
- George Konidaris
Organizations
- Brown University
- Office of Naval Research
- United States Navy