Nicholas Armstrong-Crews | Massachusetts Institute of Technology (MIT) (original) (raw)

Papers by Nicholas Armstrong-Crews

International Journal of Humanoid Robotics, Sep 1, 2008

We present our work on creating a team of two humanoid robot commentators for soccer games of tea... more We present our work on creating a team of two humanoid robot commentators for soccer games of teams of four AIBO robots. The two humanoids stand on the side lines of the field, autonomously observe the game, wirelessly listen to a "game computer controller," and coordinate their announcements with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a "puppet master" system that allows humans to intervene in a sliding autonomy manner, prompting the robots to commentate on an event if undetected. The robots process then input from these three sources, namely own and shared vision, game controller, and occasional puppet master, to recognize events which they translate into a varied set of predefined announcements. We present the behavioral architecture, the vision-based event recognition, and the game-based adaptive criteria for the selection of comments. We exemplify our development with multiple illustrative cases corresponding to different game situations. In summary, our work contributes a team of two humanoids fully executing a challenging observation, modeling, coordination, and reporting task.

Lecture Notes in Computer Science, 2022

Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are ... more Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are synergistic as they capture complementary information and work well under different environmental conditions. Fusing camera and radar data is challenging, however, as each of the sensors lacks information along a perpendicular axis, that is, depth is unknown to camera and elevation is unknown to radar. We propose the camera-radar matching network CramNet, an efficient approach to fuse the sensor readings from camera and radar in a joint 3D space. To leverage radar range measurements for better camera depth predictions, we propose a novel ray-constrained cross-attention mechanism that resolves the ambiguity in the geometric correspondences between camera features and radar features. Our method supports training with sensor modality dropout, which leads to robust 3D object detection, even when a camera or radar sensor suddenly malfunctions on a vehicle. We demonstrate the effectiveness of our fusion approach through extensive experiments on the RADIATE dataset, one of the few large-scale datasets that provide radar radio frequency imagery. A camera-only variant of our method achieves competitive performance in monocular 3D object detection on the Waymo Open Dataset.

arXiv (Cornell University), Oct 17, 2022

Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are ... more Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are synergistic as they capture complementary information and work well under different environmental conditions. Fusing camera and radar data is challenging, however, as each of the sensors lacks information along a perpendicular axis, that is, depth is unknown to camera and elevation is unknown to radar. We propose the camera-radar matching network CramNet, an efficient approach to fuse the sensor readings from camera and radar in a joint 3D space. To leverage radar range measurements for better camera depth predictions, we propose a novel ray-constrained cross-attention mechanism that resolves the ambiguity in the geometric correspondences between camera features and radar features. Our method supports training with sensor modality dropout, which leads to robust 3D object detection, even when a camera or radar sensor suddenly malfunctions on a vehicle. We demonstrate the effectiveness of our fusion approach through extensive experiments on the RADIATE dataset, one of the few large-scale datasets that provide radar radio frequency imagery. A camera-only variant of our method achieves competitive performance in monocular 3D object detection on the Waymo Open Dataset.

Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2014, 2014

A mobile robot operating in a netcentric environment can utilize offboard resources on the networ... more A mobile robot operating in a netcentric environment can utilize offboard resources on the network to improve its local perception. One such offboard resource is a world model built and maintained by other sensor systems. In this paper we present results from research into improving the performance of Deformable Parts Model object detection algorithms by using an offboard 3D world model. Experiments were run for detecting both people and cars in 2D photographs taken in an urban environment. After generating candidate object detections, a 3D world model built from airborne Light Detection and Ranging (LIDAR) and aerial photographs was used to filter out false alarm using several types of geometric reasoning. Comparison of the baseline detection performance to the performance after false alarm filtering showed a significant decrease in false alarms for a given probability of detection.

2013 IEEE International Conference on Acoustics, Speech and Signal Processing, 2013

3D point cloud registration is traditionally done by aligning to known information. This informat... more 3D point cloud registration is traditionally done by aligning to known information. This information can be extracted from semantically labeled and geo-registered 2D images, e.g. maps, satellite images, and labeled aerial photos. We propose an automated method to geo-register 3D point clouds to 2D maps by defining a normalized Hough similarity function and aligning planes (i.e., walls) in 3D point clouds to lines in 2D maps. The collective set of algorithms solves for seven degrees of freedom: three rotation parameters (including the up vector), a scale value, and three translation parameters. After transforming the 3D point cloud into a manageable 2D representation, we apply existing and novel scan-matching techniques to align both query and reference representations.

sites in remote locations for such purposes as counting fish, measuring weather, and various rese... more sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over the state, so optimal or near-optimal routing of these helicopters is paramount. Currently, the “by-eye ” technique is used, in which a route is chosen that simply looks like it would have the lowest total distance. Problem Description The well-known underlying problem at hand is the Vehicle Routing Problem (or VRP) in which a fleet of vehicles with a given capacity must make deliveries to a set of sites (Toth and Vigo 2001); this variant, however, only includes a subset of the problem specification: • A single ...

— We present our work on creating a team of two humanoid robot commentators for soccer games of t... more — We present our work on creating a team of two humanoid robot commentators for soccer games of teams of four AIBO robots. The two humanoids stand on the side lines of the field, autonomously observe the game, wirelessly listen to a “game computer controller, ” and coordinate their announcements with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a “puppet master” system that allows humans to intervene in a sliding autonomy manner, prompting the robots to commentate on an event if undetected. The robots process then input from these three sources, namely own and shared vision, game controller, and occasional puppet master, to recognize events which they translate into a varied set of predefined announcements. We present the behavioral architecture, the vision-based event recognition, and the game-based adaptive criteria for the selection of comments. We exemplify our development with multiple illustrative cases correspo...

NASA is planning to send humans and robots back to the Moon before 2020. In order for extended mi... more NASA is planning to send humans and robots back to the Moon before 2020. In order for extended missions to be productive, high quality maps of lunar terrain and resources are required. Although orbital images can provide much information, many features (local topography, resources, etc) will have to be characterized directly on the surface. To address this need, we are developing a system to perform site survey and sampling. The system includes multiple robots and humans operating in a variety of team configurations, coordinated via peer-to-peer human-robot interaction. In this paper, we present our system design and describe planned field tests. I.

— In this paper, we describe an indoor robot capable of autonomous localization and navigation th... more — In this paper, we describe an indoor robot capable of autonomous localization and navigation that serves as a companion to a human visitor. We contribute a fully-functional robot that can localize itself using static wireless nodes, manage its own uncertainty, and ask humans for help when needed. We define the indoor map as a graph embedded in the Cartesian plane with an associated matrix of learned wireless signal strengths. The localization algorithm uses the sensed wireless signals and the map constraints to generate a belief about possible locations of the robot. In addition to its own state, the robot must maintain uncertain hypotheses about the visitor’s state and the task state. It must reason about these uncertain beliefs in order to successfully execute its task, which sometimes requires querying the visitor for assistance. The robot can also interact with other humans, such as a meeting host or administrative assistant, to better manage the visit. We present illustration...

Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety ... more Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety of decision problems; unfortunately, solving for an optimal policy is computational intractable in general. In this paper, we present a set of novel search techniques for solving POMDPs approximately. We build on previous heuristic search and point-based algorithms, but improve upon them in several ways: we introduce an efficient method for approximating the convex hull of the upper bound, we expose embedded finite Markov structure in each bound, and we show how to prune aggressively while still maintaining convergence. The net result is a more targeted growth of the bound representations, leading to lower overall runtime and storage. We synthesize these contributions into a novel algorithm, which we call AAA-POMDP (Appropriately Acronymmed Algorithm). We describe its theoretical properties, including computational efficiency, and examine its performance on on standard benchmark problems...

We present a team of two humanoid robot commentators for AIBO robot soccer games. The two humanoi... more We present a team of two humanoid robot commentators for AIBO robot soccer games. The two humanoids stand by the side lines of the playing field, autonomously observe the game, wirelessly listen to a “game controller ” computer, recognize events, and select announcing actions that may require coordination with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a “Puppet Master ” control that allows humans to intervene, prompting the robots to commentate an event if previously undefined or undetected. The robots recognize events based on input from these three sources, namely own and shared vision, game controller, and occasional Puppet Master. We present the two-humanoid behavioral architecture and the vision-based event recognition, including a SIFT-based vision processing algorithm that allows for the detection of multiple similar objects, such as the identical shaped robot players. We introduce the commentating algorithm...

Abstract-We propose a new approximate algorithm, LA-JIV (Lookahead J-MDP Information Value), to s... more Abstract-We propose a new approximate algorithm, LA-JIV (Lookahead J-MDP Information Value), to solve Oracular Partially Observable Markov Decision Problems (OPOMDPs), a special type of POMDP that rather than standard observations includes an "oracle" that can be consulted for full state information at a fixed cost. We previously introduced JIV (J-MDP Information Value) to solve OPOMDPs, an heuristic algorithm that utilizes the solution of the underlying MDP and weighs the value of consulting the oracle against the value of taking a state-modifying action. While efficient, JIV will rarely find the optimal solution. In this paper, we extend JIV to include lookahead, thereby permitting arbitrarily small deviation from the optimal policy's long-term expected reward at the cost of added computation time. The depth of the lookahead is a parameter that governs this tradeoff; by iteratively increasing this depth, we provide an anytime algorithm that yields an everimproving s...

Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety ... more Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety of decision problems; unfortunately, solving for an optimal policy is computational intractable in general. In this paper, we present a set of novel search techniques for solving POMDPs approximately. We build on previous heuristic search and point-based algorithms, but improve upon them in several ways: we introduce an efficient method for approximating the convex hull of the upper bound, we expose embedded finite Markov structure in each bound, and we show how to prune aggressively while still maintaining convergence. The net result is a more targeted growth of the bound representations, leading to lower overall runtime and storage. We synthesize these contributions into a novel algorithm, which we call AAA-POMDP (Appropriately Acronymmed Algorithm). We describe its theoretical properties, including computational efficiency, and examine its performance on on standard benchmark problems...

Waves are everywhere; countless natural phenomena are composed of waves or interact with waves in... more Waves are everywhere; countless natural phenomena are composed of waves or interact with waves in some way. Light, sound, earthquakes, oceanic waves – all are examples of waves. Since waves are so ubiquitous, it is of great importance for scientists and engineers to study and model them. For example, a seismologist might want to model an earthquake’s shockwave. The wave simulator described in this document is designed to simulate such a situation, i.e., the propagation of an acoustic wave through a medium of varying density. It models the wave’s propagation in two dimensions, using a finite difference method and absorbing boundary conditions. In order to accommodate large sets of data over many time iterations and to model such data in a tractable amount of time, the simulator is parallelized using MPI. The output is displayed visually as an animation over a three-dimensional surface, in which the z-axis represents the density of the medium; the wave can be seen as a displacement in...

In this paper, we describe an indoor robot capable of autonomous localization and navigation that... more In this paper, we describe an indoor robot capable of autonomous localization and navigation that serves as a companion to a human visitor. We contribute a fully-functional robot that can localize itself using static wireless nodes, manage its own uncertainty, and ask humans for help when needed. We define the indoor map as a graph embedded in the Cartesian plane with an associated matrix of learned wireless signal strengths. The localization algorithm uses the sensed wireless signals and the map constraints to generate a belief about possible locations of the robot. In addition to its own state, the robot must maintain uncertain hypotheses about the visitor’s state and the task state. It must reason about these uncertain beliefs in order to successfully execute its task, which sometimes requires querying the visitor for assistance. The robot can also interact with other humans, such as a meeting host or administrative assistant, to better manage the visit. We present illustrations ...

The Department of Fish and Game keeps observational sites in remote locations for such purposes a... more The Department of Fish and Game keeps observational sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over the state, so optimal or near-optimal routing of these helicopters is paramount. Currently, the “by-eye” technique is used, in which a route is chosen that simply looks like it would have the lowest total distance.

Small autonomous vehicles can carry only a limited assortment of sensing and computational device... more Small autonomous vehicles can carry only a limited assortment of sensing and computational devices. One method of increasing these vehicles’ capabilities is to access offboard, networked resources. A useful resource for robotic systems would be a three‐dimensional model of the environment, continually updated and made available via netcentric applications. Researchers at Lincoln Laboratory are exploring the use of such a world model by autonomous vehicles to improve the detection of objects of interest.

PROCEEDINGS OF THE NATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE, Jul 9, 2005

Background The Department of Fish and Game keeps observational sites in remote locations for such... more Background The Department of Fish and Game keeps observational sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over ...

In this research thrust, the aim is to invert the paradigm of humans controlling (tele-operating)... more In this research thrust, the aim is to invert the paradigm of humans controlling (tele-operating) robots, and instead use autonomous robots as valuable team members - providing support by giving relevant, actionable information when appropriate. In this work, we focus on spatially-tagged cues, which can best be communicated to human team members in their local spatial context through an augmented reality display. In particular, our robots employ a custom 3D thermal infrared sensor, allowing them to detect actors that might otherwise be obscured or invisible to humans. In this fashion, the human team members have their capabilities augmented with super-human sensing.

3D point cloud registration is traditionally done by aligning to known information. This informat... more 3D point cloud registration is traditionally done by aligning to known information. This information can be extracted from semantically labeled and geo-registered 2D images, e.g. maps, satellite images, and labeled aerial photos. We propose an automated method to geo-register 3D point clouds to 2D maps by defining a normalized Hough similarity function and aligning planes (i.e., walls) in 3D point clouds to lines in 2D maps. The collective set of algorithms solves for seven degrees of freedom: three rotation parameters (including the up vector), a scale value, and three translation parameters. After transforming the 3D point cloud into a manageable 2D representation, we apply existing and novel scan-matching techniques to align both query and reference representations.

International Journal of Humanoid Robotics, Sep 1, 2008

We present our work on creating a team of two humanoid robot commentators for soccer games of tea... more We present our work on creating a team of two humanoid robot commentators for soccer games of teams of four AIBO robots. The two humanoids stand on the side lines of the field, autonomously observe the game, wirelessly listen to a "game computer controller," and coordinate their announcements with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a "puppet master" system that allows humans to intervene in a sliding autonomy manner, prompting the robots to commentate on an event if undetected. The robots process then input from these three sources, namely own and shared vision, game controller, and occasional puppet master, to recognize events which they translate into a varied set of predefined announcements. We present the behavioral architecture, the vision-based event recognition, and the game-based adaptive criteria for the selection of comments. We exemplify our development with multiple illustrative cases corresponding to different game situations. In summary, our work contributes a team of two humanoids fully executing a challenging observation, modeling, coordination, and reporting task.

Lecture Notes in Computer Science, 2022

Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are ... more Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are synergistic as they capture complementary information and work well under different environmental conditions. Fusing camera and radar data is challenging, however, as each of the sensors lacks information along a perpendicular axis, that is, depth is unknown to camera and elevation is unknown to radar. We propose the camera-radar matching network CramNet, an efficient approach to fuse the sensor readings from camera and radar in a joint 3D space. To leverage radar range measurements for better camera depth predictions, we propose a novel ray-constrained cross-attention mechanism that resolves the ambiguity in the geometric correspondences between camera features and radar features. Our method supports training with sensor modality dropout, which leads to robust 3D object detection, even when a camera or radar sensor suddenly malfunctions on a vehicle. We demonstrate the effectiveness of our fusion approach through extensive experiments on the RADIATE dataset, one of the few large-scale datasets that provide radar radio frequency imagery. A camera-only variant of our method achieves competitive performance in monocular 3D object detection on the Waymo Open Dataset.

arXiv (Cornell University), Oct 17, 2022

Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are ... more Robust 3D object detection is critical for safe autonomous driving. Camera and radar sensors are synergistic as they capture complementary information and work well under different environmental conditions. Fusing camera and radar data is challenging, however, as each of the sensors lacks information along a perpendicular axis, that is, depth is unknown to camera and elevation is unknown to radar. We propose the camera-radar matching network CramNet, an efficient approach to fuse the sensor readings from camera and radar in a joint 3D space. To leverage radar range measurements for better camera depth predictions, we propose a novel ray-constrained cross-attention mechanism that resolves the ambiguity in the geometric correspondences between camera features and radar features. Our method supports training with sensor modality dropout, which leads to robust 3D object detection, even when a camera or radar sensor suddenly malfunctions on a vehicle. We demonstrate the effectiveness of our fusion approach through extensive experiments on the RADIATE dataset, one of the few large-scale datasets that provide radar radio frequency imagery. A camera-only variant of our method achieves competitive performance in monocular 3D object detection on the Waymo Open Dataset.

Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2014, 2014

A mobile robot operating in a netcentric environment can utilize offboard resources on the networ... more A mobile robot operating in a netcentric environment can utilize offboard resources on the network to improve its local perception. One such offboard resource is a world model built and maintained by other sensor systems. In this paper we present results from research into improving the performance of Deformable Parts Model object detection algorithms by using an offboard 3D world model. Experiments were run for detecting both people and cars in 2D photographs taken in an urban environment. After generating candidate object detections, a 3D world model built from airborne Light Detection and Ranging (LIDAR) and aerial photographs was used to filter out false alarm using several types of geometric reasoning. Comparison of the baseline detection performance to the performance after false alarm filtering showed a significant decrease in false alarms for a given probability of detection.

2013 IEEE International Conference on Acoustics, Speech and Signal Processing, 2013

3D point cloud registration is traditionally done by aligning to known information. This informat... more 3D point cloud registration is traditionally done by aligning to known information. This information can be extracted from semantically labeled and geo-registered 2D images, e.g. maps, satellite images, and labeled aerial photos. We propose an automated method to geo-register 3D point clouds to 2D maps by defining a normalized Hough similarity function and aligning planes (i.e., walls) in 3D point clouds to lines in 2D maps. The collective set of algorithms solves for seven degrees of freedom: three rotation parameters (including the up vector), a scale value, and three translation parameters. After transforming the 3D point cloud into a manageable 2D representation, we apply existing and novel scan-matching techniques to align both query and reference representations.

sites in remote locations for such purposes as counting fish, measuring weather, and various rese... more sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over the state, so optimal or near-optimal routing of these helicopters is paramount. Currently, the “by-eye ” technique is used, in which a route is chosen that simply looks like it would have the lowest total distance. Problem Description The well-known underlying problem at hand is the Vehicle Routing Problem (or VRP) in which a fleet of vehicles with a given capacity must make deliveries to a set of sites (Toth and Vigo 2001); this variant, however, only includes a subset of the problem specification: • A single ...

— We present our work on creating a team of two humanoid robot commentators for soccer games of t... more — We present our work on creating a team of two humanoid robot commentators for soccer games of teams of four AIBO robots. The two humanoids stand on the side lines of the field, autonomously observe the game, wirelessly listen to a “game computer controller, ” and coordinate their announcements with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a “puppet master” system that allows humans to intervene in a sliding autonomy manner, prompting the robots to commentate on an event if undetected. The robots process then input from these three sources, namely own and shared vision, game controller, and occasional puppet master, to recognize events which they translate into a varied set of predefined announcements. We present the behavioral architecture, the vision-based event recognition, and the game-based adaptive criteria for the selection of comments. We exemplify our development with multiple illustrative cases correspo...

NASA is planning to send humans and robots back to the Moon before 2020. In order for extended mi... more NASA is planning to send humans and robots back to the Moon before 2020. In order for extended missions to be productive, high quality maps of lunar terrain and resources are required. Although orbital images can provide much information, many features (local topography, resources, etc) will have to be characterized directly on the surface. To address this need, we are developing a system to perform site survey and sampling. The system includes multiple robots and humans operating in a variety of team configurations, coordinated via peer-to-peer human-robot interaction. In this paper, we present our system design and describe planned field tests. I.

— In this paper, we describe an indoor robot capable of autonomous localization and navigation th... more — In this paper, we describe an indoor robot capable of autonomous localization and navigation that serves as a companion to a human visitor. We contribute a fully-functional robot that can localize itself using static wireless nodes, manage its own uncertainty, and ask humans for help when needed. We define the indoor map as a graph embedded in the Cartesian plane with an associated matrix of learned wireless signal strengths. The localization algorithm uses the sensed wireless signals and the map constraints to generate a belief about possible locations of the robot. In addition to its own state, the robot must maintain uncertain hypotheses about the visitor’s state and the task state. It must reason about these uncertain beliefs in order to successfully execute its task, which sometimes requires querying the visitor for assistance. The robot can also interact with other humans, such as a meeting host or administrative assistant, to better manage the visit. We present illustration...

Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety ... more Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety of decision problems; unfortunately, solving for an optimal policy is computational intractable in general. In this paper, we present a set of novel search techniques for solving POMDPs approximately. We build on previous heuristic search and point-based algorithms, but improve upon them in several ways: we introduce an efficient method for approximating the convex hull of the upper bound, we expose embedded finite Markov structure in each bound, and we show how to prune aggressively while still maintaining convergence. The net result is a more targeted growth of the bound representations, leading to lower overall runtime and storage. We synthesize these contributions into a novel algorithm, which we call AAA-POMDP (Appropriately Acronymmed Algorithm). We describe its theoretical properties, including computational efficiency, and examine its performance on on standard benchmark problems...

We present a team of two humanoid robot commentators for AIBO robot soccer games. The two humanoi... more We present a team of two humanoid robot commentators for AIBO robot soccer games. The two humanoids stand by the side lines of the playing field, autonomously observe the game, wirelessly listen to a “game controller ” computer, recognize events, and select announcing actions that may require coordination with each other. Given the large degree of uncertainty and dynamics of the robot soccer games, we further introduce a “Puppet Master ” control that allows humans to intervene, prompting the robots to commentate an event if previously undefined or undetected. The robots recognize events based on input from these three sources, namely own and shared vision, game controller, and occasional Puppet Master. We present the two-humanoid behavioral architecture and the vision-based event recognition, including a SIFT-based vision processing algorithm that allows for the detection of multiple similar objects, such as the identical shaped robot players. We introduce the commentating algorithm...

Abstract-We propose a new approximate algorithm, LA-JIV (Lookahead J-MDP Information Value), to s... more Abstract-We propose a new approximate algorithm, LA-JIV (Lookahead J-MDP Information Value), to solve Oracular Partially Observable Markov Decision Problems (OPOMDPs), a special type of POMDP that rather than standard observations includes an "oracle" that can be consulted for full state information at a fixed cost. We previously introduced JIV (J-MDP Information Value) to solve OPOMDPs, an heuristic algorithm that utilizes the solution of the underlying MDP and weighs the value of consulting the oracle against the value of taking a state-modifying action. While efficient, JIV will rarely find the optimal solution. In this paper, we extend JIV to include lookahead, thereby permitting arbitrarily small deviation from the optimal policy's long-term expected reward at the cost of added computation time. The depth of the lookahead is a parameter that governs this tradeoff; by iteratively increasing this depth, we provide an anytime algorithm that yields an everimproving s...

Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety ... more Partially Observable Markov Decision Processes, or POMDPs, are useful for representing a variety of decision problems; unfortunately, solving for an optimal policy is computational intractable in general. In this paper, we present a set of novel search techniques for solving POMDPs approximately. We build on previous heuristic search and point-based algorithms, but improve upon them in several ways: we introduce an efficient method for approximating the convex hull of the upper bound, we expose embedded finite Markov structure in each bound, and we show how to prune aggressively while still maintaining convergence. The net result is a more targeted growth of the bound representations, leading to lower overall runtime and storage. We synthesize these contributions into a novel algorithm, which we call AAA-POMDP (Appropriately Acronymmed Algorithm). We describe its theoretical properties, including computational efficiency, and examine its performance on on standard benchmark problems...

Waves are everywhere; countless natural phenomena are composed of waves or interact with waves in... more Waves are everywhere; countless natural phenomena are composed of waves or interact with waves in some way. Light, sound, earthquakes, oceanic waves – all are examples of waves. Since waves are so ubiquitous, it is of great importance for scientists and engineers to study and model them. For example, a seismologist might want to model an earthquake’s shockwave. The wave simulator described in this document is designed to simulate such a situation, i.e., the propagation of an acoustic wave through a medium of varying density. It models the wave’s propagation in two dimensions, using a finite difference method and absorbing boundary conditions. In order to accommodate large sets of data over many time iterations and to model such data in a tractable amount of time, the simulator is parallelized using MPI. The output is displayed visually as an animation over a three-dimensional surface, in which the z-axis represents the density of the medium; the wave can be seen as a displacement in...

In this paper, we describe an indoor robot capable of autonomous localization and navigation that... more In this paper, we describe an indoor robot capable of autonomous localization and navigation that serves as a companion to a human visitor. We contribute a fully-functional robot that can localize itself using static wireless nodes, manage its own uncertainty, and ask humans for help when needed. We define the indoor map as a graph embedded in the Cartesian plane with an associated matrix of learned wireless signal strengths. The localization algorithm uses the sensed wireless signals and the map constraints to generate a belief about possible locations of the robot. In addition to its own state, the robot must maintain uncertain hypotheses about the visitor’s state and the task state. It must reason about these uncertain beliefs in order to successfully execute its task, which sometimes requires querying the visitor for assistance. The robot can also interact with other humans, such as a meeting host or administrative assistant, to better manage the visit. We present illustrations ...

The Department of Fish and Game keeps observational sites in remote locations for such purposes a... more The Department of Fish and Game keeps observational sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over the state, so optimal or near-optimal routing of these helicopters is paramount. Currently, the “by-eye” technique is used, in which a route is chosen that simply looks like it would have the lowest total distance.

Small autonomous vehicles can carry only a limited assortment of sensing and computational device... more Small autonomous vehicles can carry only a limited assortment of sensing and computational devices. One method of increasing these vehicles’ capabilities is to access offboard, networked resources. A useful resource for robotic systems would be a three‐dimensional model of the environment, continually updated and made available via netcentric applications. Researchers at Lincoln Laboratory are exploring the use of such a world model by autonomous vehicles to improve the detection of objects of interest.

PROCEEDINGS OF THE NATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE, Jul 9, 2005

Background The Department of Fish and Game keeps observational sites in remote locations for such... more Background The Department of Fish and Game keeps observational sites in remote locations for such purposes as counting fish, measuring weather, and various research projects. In Alaska, fishing and natural tourism are vital to the economy of the state, making these measurements all the more important. However, due to large geographic spread with little transportation infrastructure (none, between many villages), helicopters must be used to visit these sites. It is extremely expensive, primarily in fuel costs, to fly helicopters all over ...

In this research thrust, the aim is to invert the paradigm of humans controlling (tele-operating)... more In this research thrust, the aim is to invert the paradigm of humans controlling (tele-operating) robots, and instead use autonomous robots as valuable team members - providing support by giving relevant, actionable information when appropriate. In this work, we focus on spatially-tagged cues, which can best be communicated to human team members in their local spatial context through an augmented reality display. In particular, our robots employ a custom 3D thermal infrared sensor, allowing them to detect actors that might otherwise be obscured or invisible to humans. In this fashion, the human team members have their capabilities augmented with super-human sensing.

3D point cloud registration is traditionally done by aligning to known information. This informat... more 3D point cloud registration is traditionally done by aligning to known information. This information can be extracted from semantically labeled and geo-registered 2D images, e.g. maps, satellite images, and labeled aerial photos. We propose an automated method to geo-register 3D point clouds to 2D maps by defining a normalized Hough similarity function and aligning planes (i.e., walls) in 3D point clouds to lines in 2D maps. The collective set of algorithms solves for seven degrees of freedom: three rotation parameters (including the up vector), a scale value, and three translation parameters. After transforming the 3D point cloud into a manageable 2D representation, we apply existing and novel scan-matching techniques to align both query and reference representations.