About

Matthew Elwin is an Associate Professor of Instruction in Mechanical Engineering at Northwestern University and serves as the Director of the Master of Science in Robotics Program. His educational background includes a PhD and MS in mechanical engineering from Northwestern University, as well as a BE with an emphasis in control systems from Dartmouth College, and an AB in engineering sciences from Dartmouth College. His research interests encompass swarm robotics, cooperative robot manipulation and dexterity, robotic navigation, and mechatronics. Elwin has contributed to the field through various publications on topics such as cooperative payload estimation, human-multirobot collaborative mobile manipulation, distributed environmental monitoring with finite element robots, fault detection in dynamic consensus, and environmental estimation with distributed finite element agents. In the classroom, he teaches robotics courses, advises robotics projects, and develops curriculum for the MS in Robotics program.

Research topics

• Artificial Intelligence
• Computer Science
• Physical medicine and rehabilitation
• Telecommunications
• Geometry
• Acoustics
• Mathematics
• Medicine
• Computer network
• Mathematical analysis
• Engineering
• Control engineering
• Human–computer interaction
• Physics

Selected publications

• Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

  ArXiv.org · 2026-03-02

  articleOpen access

  Post-stroke rehabilitation is often necessary for patients to regain proper walking gait. However, the typical therapy process can be exhausting and physically demanding for therapists, potentially reducing therapy intensity, duration, and consistency over time. We propose a Patient-Therapist Force Field (PTFF) to visualize therapist responses to patient kinematics and a Synthetic Therapist (ST) machine learning model to support the therapist in dyadic robot-mediated physical interaction therapy. The first encodes patient and therapist stride kinematics into a shared low-dimensional latent manifold using a Variational Autoencoder (VAE) and models their interaction through a Gaussian Mixture Model (GMM), which learns a probabilistic vector field mapping patient latent states to therapist responses. This representation visualizes patient-therapist interaction dynamics to inform therapy strategies and robot controller design. The latter is implemented as a Long Short-Term Memory (LSTM) network trained on patient-therapist interaction data to predict therapist-applied joint torques from patient kinematics. Trained and validated using leave-one-out cross-validation across eight post-stroke patients, the model was integrated into a ROS-based exoskeleton controller to generate real-time torque assistance based on predicted therapist responses. Offline results and preliminary testing indicate the potential of their use as an alternative approach to post-stroke exoskeleton therapy. The PTFF provides understanding of the therapist's actions while the ST frees the human therapist from the exoskeleton, allowing them to continuously monitor the patient's nuanced condition.

  PublisherOA PDF

• Dead Zone Bilateral Control for High Performance Robotic Teleoperation

  IEEE Robotics and Automation Letters · 2026-03-19

  article

  Autonomous robot policies are commonly trained using demonstration data acquired via robotic teleoperation, a process which can be time-intensive and physically demanding for human operators. Bilateral control can speed up robotic teleoperation by allowing the operator to feel the forces experienced by the remote manipulator, but it can also transmit undesirable forces, such as friction and damping, back to the operator. Here, we present dead zone bilateral control, an improvement to the conventional position-position bilateral control scheme that prevents the transmission of unwanted forces during free-space motion. We implement our controller on a custom 2-degree-of-freedom teleoperation device and show that it reduces the energy required for free-space motion by approximately 50%. During a user study in which 16 participants performed a peg rolling task, our controller reduced completion times by an average of 25% compared to the standard position-position bilateral controller. The results suggest that dead zone bilateral control can expedite the collection of teleoperated task demonstrations, allowing researchers to gather larger datasets for training autonomous robot policies.

  PublisherDOI

• Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

  arXiv (Cornell University) · 2026-03-02

  preprintOpen access

  Post-stroke rehabilitation is often necessary for patients to regain proper walking gait. However, the typical therapy process can be exhausting and physically demanding for therapists, potentially reducing therapy intensity, duration, and consistency over time. We propose a Patient-Therapist Force Field (PTFF) to visualize therapist responses to patient kinematics and a Synthetic Therapist (ST) machine learning model to support the therapist in dyadic robot-mediated physical interaction therapy. The first encodes patient and therapist stride kinematics into a shared low-dimensional latent manifold using a Variational Autoencoder (VAE) and models their interaction through a Gaussian Mixture Model (GMM), which learns a probabilistic vector field mapping patient latent states to therapist responses. This representation visualizes patient-therapist interaction dynamics to inform therapy strategies and robot controller design. The latter is implemented as a Long Short-Term Memory (LSTM) network trained on patient-therapist interaction data to predict therapist-applied joint torques from patient kinematics. Trained and validated using leave-one-out cross-validation across eight post-stroke patients, the model was integrated into a ROS-based exoskeleton controller to generate real-time torque assistance based on predicted therapist responses. Offline results and preliminary testing indicate the potential of their use as an alternative approach to post-stroke exoskeleton therapy. The PTFF provides understanding of the therapist's actions while the ST frees the human therapist from the exoskeleton, allowing them to continuously monitor the patient's nuanced condition.

  PublisherDOI

• Cooperative Payload Estimation by a Team of Mocobots

  ArXiv.org · 2025-02-07

  preprintOpen access

  For high-performance autonomous manipulation of a payload by a mobile manipulator team, or for collaborative manipulation with the human, robots should be able to discover where other robots are attached to the payload, as well as the payload's mass and inertial properties. In this paper, we describe a method for the robots to autonomously discover this information. The robots cooperatively manipulate the payload, and the twist, twist derivative, and wrench data at their grasp frames are used to estimate the transformation matrices between the grasp frames, the location of the payload's center of mass, and the payload's inertia matrix. The method is validated experimentally with a team of three mobile cobots, or mocobots.

  PublisherOA PDFDOI

• Design and Control of an Actuated Prosthetic Elbow to Restore Arm Swing for Persons with Upper Limb Absence

  2025-05-12

  article

  Dynamic balance in human locomotion relies on the regulation of whole-body angular momentum. Asymmetry in upper extremity body mass and motion that occurs from upper limb absence disrupts this regulation. Conventional transhumeral prostheses do not restore controlled arm swing, leaving this asymmetry unresolved. This project aimed to develop an actuated prosthetic elbow that mimics natural elbow rotations to restore arm swing during walking. Two prototypes with contrasting designs were fabricated and tested. Prototype 1, featuring a direct drive motor, tracked arm swing trajectories based on able-bodied data with less than 5 % position error. It enabled evaluation of predicted elbow torque requirements, showing that a double-pendulum dynamic model overestimates torque needs. This finding supports using smaller, lighter direct drive motors for natural arm swing. While Prototype 1 performed well, its motor was too large for practical integration. Prototype 2 successfully fit the drive and control system into a standard prosthetic envelope for enhanced applicability. Testing revealed that gearbox friction hindered control, and successful operation required synchronizing elbow movement with natural arm swing frequency. Despite using the same control system, both prototypes produced notably different outcomes. These results highlight potential for future designs incorporating a small direct drive motor to match Prototype 1's performance while maintaining Prototype 2's compact form.

  PublisherDOI

• Cooperative Payload Estimation by a Team of Mocobots

  IEEE Robotics and Automation Letters · 2025-08-11 · 1 citations

  article

  For high-performance autonomous manipulation of a payload by a mobile manipulator team, or for collaborative manipulation with the human, robots should be able to discover where other robots are attached to the payload, as well as the payload's mass and inertial properties. In this paper, we describe a method for the robots to autonomously discover this information. The robots cooperatively manipulate the payload, and the twist, twist derivative, and wrench data at their grasp frames are used to estimate the transformation matrices between the grasp frames, the location of the payload's center of mass, and the payload's inertia matrix. The method is validated experimentally with a team of three mobile cobots, or mocobots.

  PublisherDOI

• Self-Healing Distributed Swarm Formation Control Using Image Moments

  IEEE Robotics and Automation Letters · 2024-05-15 · 2 citations

  article

  Human-swarm interaction is facilitated by a low-dimensional encoding of the swarm formation, independent of the (possibly large) number of robots. We propose using image moments to encode two-dimensional formations of robots. Each robot knows its pose and the desired formation moments, and simultaneously estimates the current moments of the entire swarm while controlling its motion to better achieve the desired group moments. The estimator is a distributed optimization, requiring no centralized processing, and self-healing, meaning that the process is robust to initialization errors, packet drops, and robots being added to or removed from the swarm. Our experimental results with a swarm of 50 robots, suffering nearly 50% packet loss, show that distributed estimation and control of image moments effectively achieves desired swarm formations.

  PublisherDOI

• Self-Healing Distributed Swarm Formation Control Using Image Moments

  arXiv (Cornell University) · 2023-12-12

  preprintOpen access

  Human-swarm interaction is facilitated by a low-dimensional encoding of the swarm formation, independent of the (possibly large) number of robots. We propose using image moments to encode two-dimensional formations of robots. Each robot knows its pose and the desired formation moments, and simultaneously estimates the current moments of the entire swarm while controlling its motion to better achieve the desired group moments. The estimator is a distributed optimization, requiring no centralized processing, and self-healing, meaning that the process is robust to initialization errors, packet drops, and robots being added to or removed from the swarm. Our experimental results with a swarm of 50 robots, suffering nearly 50% packet loss, show that distributed estimation and control of image moments effectively achieves desired swarm formations.

  PublisherOA PDFDOI

• Fabrication of Quasi-Vertical GaN-On-SiC Trench MOSFETs

  Key engineering materials · 2023-05-19

  articleOpen accessSenior author

  We demonstrate quasi-vertical GaN MOSFETs fabricated on SiC substrates. The GaN epitaxial layers were grown via MOCVD on 100 mm 4H-SiC wafers, with the device structure consisting of a 2.5 μm drift layer and a Mg doped p-GaN body. The fabricated transistors exhibit normally-off characteristics, with low off-state leakage behavior and an on/off ratio of over . The specific on-resistance was measured to be which compares favorably to devices fabricated on other foreign substrates. Our results demonstrate an alternative substrate for realizing vertical GaN devices, which potentially offers better material quality and thermal properties compared with other foreign substrate choices.

  PublisherOA PDFDOI

• Semi-supervised adaptation of upper-limb myoelectric pattern recognition prosthesis control through virtual gameplay

  2023-04-24

  article

  Advanced machine learning algorithms can adapt to variation in new data inputs. Such adaptive algorithms have been employed on myoelectric pattern recognition control systems to improve upper-limb prosthesis performance. When training their control system, prosthesis users typically attempt to make consistent and repeatable muscle contractions. However, minimizing input data variation does not always resemble realistic usage scenarios as several factors (muscle fatigue, limb position, electrode shift, etc.) can contribute to changes in the characteristics of the muscle signals that could lead to poor controller performance. While it may be difficult to account for all the possible variation, prosthesis users may benefit from training that better mimics real-life prosthesis use. This paper investigates the use of virtual games, developed for practicing specific aspects of myoelectric prosthesis control, to adapt a linear discriminant analysis (LDA) model in a semi-supervised manner. Results from offline analysis of virtual game data collected across two weeks showed that classification error rates were better for 7 out of 10 prosthesis users when applying an adaptive LDA model compared to a traditional non-adaptive LDA model. We also compare these results to an alternative model in which we apply a heuristic set of rules to identify and relabel “misclassified” predicted outputs during virtual game play before evaluating the classification performance of an adaptive LDA classifier with re-labeled inputs. Virtual games are a promising clinical tool which can be applied to better learn the user's control preferences under simulated use conditions. Further development of this work could impact daily prosthesis use and performance for those who use myoelectric pattern recognition-controlled prostheses.

  PublisherDOI

Frequent coauthors

• Kevin Lynch

  46 shared

• Nicholas Marchuk

  29 shared

• Randy A. Freeman

  Northwestern University

  15 shared

• Julius P. A. Dewald

  Northwestern University

  4 shared

• Petar Igić

  Coventry University

  3 shared

• Jemin George

  3 shared

• Ian Abraham

  2 shared

• Todd Murphey

  Northwestern University

  2 shared