Overview
The robots of the future may be called upon to search for a variety of objects and to do so in unfamiliar territory. An established method called Partially Observable Markov Decision Processes (POMDPs) allows a robot to make decisions in the correct order even faced with uncertain conditions.
However, these POMDPs struggle to keep pace as the number of places the machine must search or the number of items it is looking for increases. Tellex’s team has invented a new version of this idea called Multi-Object Search OO-POMDP (MOS OO-POMDP) that can more easily scale to large domains.
Market Opportunity
Imagine a search-and-rescue robot searching for survivors who might be spread across the site of a disaster, or a home assistant bot asked to find all the toys scattered across the living room as part of its cleaning tasks. These are examples of the kind of open-ended search that robots of the future may be tasked with.
POMDPs are an appealing approach for modeling how a robot might approach and execute such a search, because they can capture the discrete tasks a robot would use in a partially observable, uncertain environment: observe, predict, and act. A POMDP tries to model a robot’s current and future “beliefs” about what is or might be true as it searches an unknown space. The problem is that those “belief” spaces grow exponentially with the number of objects sought. A new approach is needed to ease the computational burden.
Innovation and Meaningful Advantages
Tellex created an object-oriented OO-POMDP to represent the robot’s tasks as it conducts a search under uncertain conditions. The state, transition, and observation spaces are represented within the model in terms of classes and objects. In this way, Tellex can factor the belief into object distributions, which allows the belief space size to scale linearly—rather than exponentially—as the number of objects increases.
Crucially, Tellex has also demonstrated the ability to use natural language to update a robot’s belief space. For example, a future home assistant robot might be asked to, “find the mugs in the library, living room, or kitchen,” which grounds the robot’s expectations of uncertainty from the outset. OO-POMDPs make it simple to express observation models involving objects, as in this example. Additionally, belief updates can be easily restricted to involve only the relevant objects.
Collaboration Opportunity
We are seeking a licensing opportunity for this innovative technology.
Principal Investigator
Stefanie Tellex, PhD
Associate Professor of Computer Science; Associate Professor of Engineering
Brown University
IP Information
US Utility US20210347046A1, Issued February 13, 2024
Contact
Brian Demers
Director of Business Development, School of Engineering and Physics
Brown Tech ID: 2582