Informational seminar: Trajectory Planning and Constraint Guarantees with Neural Networks
Abstract: This seminar will first study the problem of trajectory
planning in environments with static obstacles, dynamic obstacles, and other
planning agents. It will present the MINVO basis, a polynomial basis that
substantially reduces the conservativeness in the obstacle avoidance
constraints. Leveraging this MINVO basis, it will then explain a
computationally efficient algorithm called MADER to avoid dynamic obstacles and
other agents in a decentralized and asynchronous way. This framework will then
be extended to include perception-awareness, using both model-based and
learning-based approaches (PANTHER and Deep-PANTHER). The seminar will end with
the presentation of a novel algorithm called RAYEN that enables the imposition
of hard convex constraints on neural networks.
Bio: Jesús Tordesillas received his Ph.D. from MIT in 2022 as a
member of the Aerospace Controls Laboratory under the supervision of Prof.
Jonathan P. How. He also did an internship at the NASA-Jet Propulsion
Laboratory working for the DARPA Subterranean Challenge. After his Ph.D., he
became a postdoctoral associate at MIT and later at ETH Zurich in the Robotic
Systems Laboratory led by Prof. Marco Hutter. He is currently an Assistant
Professor at ICAI, Comillas Pontifical University. His research interests
include trajectory planning in dynamic environments, optimization, and
constrained deep learning. More information about his research can be found at https://jtorde.github.io/
Abstract: This seminar will first study the problem of trajectory
planning in environments with static obstacles, dynamic obstacles, and other
planning agents. It will present the MINVO basis, a polynomial basis that
substantially reduces the conservativeness in the obstacle avoidance
constraints. Leveraging this MINVO basis, it will then explain a
computationally efficient algorithm called MADER to avoid dynamic obstacles and
other agents in a decentralized and asynchronous way. This framework will then
be extended to include perception-awareness, using both model-based and
learning-based approaches (PANTHER and Deep-PANTHER). The seminar will end with
the presentation of a novel algorithm called RAYEN that enables the imposition
of hard convex constraints on neural networks.
Bio: Jesús Tordesillas received his Ph.D. from MIT in 2022 as a
member of the Aerospace Controls Laboratory under the supervision of Prof.
Jonathan P. How. He also did an internship at the NASA-Jet Propulsion
Laboratory working for the DARPA Subterranean Challenge. After his Ph.D., he
became a postdoctoral associate at MIT and later at ETH Zurich in the Robotic
Systems Laboratory led by Prof. Marco Hutter. He is currently an Assistant
Professor at ICAI, Comillas Pontifical University. His research interests
include trajectory planning in dynamic environments, optimization, and
constrained deep learning. More information about his research can be found at https://jtorde.github.io/