Tesshu Fujinami

University of Pennsylvania, PhD Student

Tesshu Fujinami

Simplicity is prerequisite for reliability.

Edsger W. Dijkstra, How do we tell truths that might hurt?, 1975
One controller, many systems.

Research

My goal is to bridge the gap between robot learning in practice and theory of learning for control, toward control systems that are reliable, robust, and adaptable. My interests lie at the intersection of control theory, reinforcement learning, and policy optimization.

Publications

  1. University of Pennsylvania: Domain Randomization theory for LQR
  2. University of Tokyo: Autonomous Drone Control
  3. Extracurricular: Autonomous Car Racing

University of Pennsylvania

  1. Robust in Training, Adaptable at Deployment: Lessons from Domain-Randomized LQR

    Tesshu Fujinami, George J. Pappas

    Sim2Real2Field Workshop, Conference on Robot Learning (CoRL), 2026

    Triangle relating domain randomization, adaptation, and forgetting.
    Broader randomization speeds adaptation and reduces forgetting, but makes optimization harder, and faster adaptation itself trades off against forgetting.
  2. Policy Gradient for LQR with Domain Randomization

    Tesshu Fujinami, Bruce D. Lee, Nikolai Matni, George J. Pappas

    IEEE Conference on Decision and Control (CDC), 2025

    Domain-randomized LQR cost versus iterations for 10, 20, and 500 sampled systems.
    Policy gradient provably converges on the domain-randomized LQR objective. Sampling more systems (larger M) brings the solution closer to the true randomized cost.
  3. Policy Gradient over History-Dependent Policy Classes for LQR with Domain Randomization

    Tesshu Fujinami, Bruce D. Lee, Anastasios Tsiamis, Nikolai Matni, George J. Pappas

    IEEE Conference on Decision and Control (CDC), 2026

    LQR cost, controller parameters, and discount factor across policy-gradient iterations, with phase boundaries.
    Growing the controller’s memory one phase at a time, while annealing the discount factor to 1, lets policy gradient stabilize system collections that no static feedback gain can.
  4. Domain Randomization is Sample Efficient for Linear Quadratic Control

    Tesshu Fujinami, Bruce D. Lee, Nikolai Matni, George J. Pappas

    Learning for Dynamics and Control Conference (L4DC), 2025

    Oral presentation (16/120)

    Control cost versus amount of data for robust control, domain randomization, and certainty equivalence.
    With little data, domain randomization beats certainty equivalence. As data grows, it matches certainty equivalence’s optimal rate, while robust control stays safe but becomes conservative.

Back to outline

University of Tokyo

  1. Gaussian Process-based Visual Pursuit Control with Automatic Gain Tuning

    Tesshu Fujinami, Junya Yamauchi, Marco Omainska, Masayuki Fujita

    IEEE Conference on Control Technology and Applications (CCTA), 2022

  1. A Control Barrier Function Approach for Observer-based Visually Safe Pursuit Control with Spherical Obstacles

    Tesshu Fujinami, Junya Yamauchi, Riku Funada, Masayuki Fujita

    IFAC World Congress, 2023

  1. Visual Pursuit with Switched Motion Estimation and Rigid Body Gaussian Processes

    Marco Omainska, Junya Yamauchi, Tesshu Fujinami, Masayuki Fujita

    Transactions of the Institute of Systems, Control and Information Engineers, 2023

    Best paper award

  2. Visual Safety of Robot Networks based on Control Barrier Functions with Spherical Obstacles

    Junya Yamauchi, Reita Maeno, Tesshu Fujinami, Marco Omainska, Masayuki Fujita

    SICE Annual Conference, 2023

  3. Distributed Gaussian Process Based Cooperative Visual Pursuit Control for Drone Networks

    Makoto Saito, Junya Yamauchi, Tesshu Fujinami, Marco Omainska, Masayuki Fujita

    IFAC Symposium on Mechatronic Systems, 2022

Back to outline

Extracurricular

F1TENTH (RoboRacer) Autonomous Racing

Tesshu Fujinami, Henry Huang, Taiki Nakano, Makoto Saito

5th place at Autonomous Grand Prix in ICRA 2024

Back to outline