Rihan Aaron D'Silva

Hello!

I am a Master's student in Aerospace Engineering at the University of Illinois Urbana-Champaign. My research interests are at the intersection of Control theory, Learning, Optimization for Robotics. I am interested in foundations for end-to-end control architectures that offer guarantees for safety, robustness and continual learning in changing uncertain environments. Currently, I am working under Prof. Hiroyasu Tsukamoto towards uncertainty quantification and safety certificates for nonlinear motion planning.

Previously, I completed my undergraduate studies at IIT Bombay, where I worked with Prof. Debasish Chatterjee on Numerical approximation of Optimal Control trajectories and at Team Aerove on challenging Aerial robotics problems.

Email / LinkedIn / GitHub / CV

Research Experience

Conformally Robust Chance-Constrained Trajectory Optimization and Control of Non-Gaussian Stochastic Systems. [Video]

This work aims to design distribution-free confidence sets for closed-loop nonlinear systems under additive stochastic noise using neural contraction metrics. Via suitable constraint tightening, we obtain marginal chance-constraint satisfaction guarantees while accounting for the coverage gap with the unknown test distribution.

Uniform Error Guarantees and Mesh Refinement for Direct Solutions to Constrained Nonlinear Optimal Control problems.

A new transcription, change point localization, and mesh refinement scheme is presented for direct optimization-based solutions and for uniform approximation (as opposed to typical L2 approximation) of optimal control trajectories associated with a class of nonlinear constrained optimal control problems. This method proves to better capture singular optimal control trajectories, when compared to existing state of art methods. Check out software QuITO v.2 for the implementation.

Reinforcement learning assisted Model Predictive control

Reinforcment learning (RL) strategies often remain impractical in process control due to long and expensive training and absence of safety guarantees. We investigate a RL supervised MPC approach that uses a closed-loop re-identification procedure to correct for plant-model mismatch, with dynamic tuning of MPC parameters using RL updates.

Projects

AUVSI International Aerial Robotics Competition 2021-22 [Project Page, Code, Video]

Objective: To autonomously perform aerial replacement of a 2kg heavy module attached to a mobile platform while operating over long distances and mission time constraints.

Impact - Awarded special mention for hardware implementation of a mothership-daughtership configuration

Terrain relative optical navigation [Code]

Existing lunar navigation technology suffer from poor performance in pinpoint lunar landing missions. This work employs a feature detection approach to develop a high-accuracy visual-based positioning algorithm for inertial state estimation during lunar descent.

IMechE Unmanned Aerial System (UAS) Competition 2023 [Code]

Design and developed UAV for a long range payload deployment mission with real-time path planning and optimization under mission time and battery consumption constraints

ICUAS Competition 2022 [Code]

Develop an autonomous fire-fighting quadcopter capable of navigating safely in indoor environments using VFH and a custom dynamic payload throw maneuver.

Control for Quadcopter trajectory tracking [Code]

Quadrotor maneuvering in three dimensions in a tightly constrained setting is a compelling practical problem. We employ time optimal minimum snap polynomial spline trajectories subject to obstacle avoidance and actuator constraints, that are tracked using robust (upto some degree) reference tracking methods such as PID, Backstepping and Sliding mode controllers. We also develop pipelines and test algorithms on Pixhawk Gazebo MAVROS and hardware.