Research & technical notes

Collision-aware control for autonomous vehicles

My graduate research explores how an autonomous vehicle controller can track a path and avoid obstacles in normal operation, then transition to damage-aware decision-making when a collision becomes unavoidable.

Working research overview — not a formal publication

Research direction

Autonomous driving controllers are normally evaluated on path-following accuracy, obstacle avoidance, comfort, and constraint satisfaction. My work adds a second regime: if avoidance is no longer physically feasible, the controller should choose the remaining action that minimizes the severity of the resulting collision.

Control architecture

The baseline controller is structured around model predictive control. The vehicle model, actuator limits, road boundaries, and obstacle constraints are represented explicitly so the optimization can reason about feasible future motion over a finite horizon.

Operational objectivePath tracking + obstacle avoidance + control smoothnessImminent-collision objectiveReduce collision severity subject to real-time and physical constraints

Development path

The project is deliberately being developed in stages. MATLAB/Simulink is used first to establish the baseline formulation, benchmark scenarios, and controller tradeoffs. The optimization layer is then planned to migrate to Python, with CasADi considered for symbolic model and constraint construction and acados considered for fast real-time NMPC. For higher-fidelity vehicle-dynamics and collision validation, the simulation stack may extend to Project Chrono while retaining MuJoCo for rapid prototyping and controlled contact experiments.

BaselineMATLAB / SimulinkReal-time migrationPython + planned CasADi / acados evaluationHigher-fidelity validationPlanned Project Chrono integration

Why computation time matters

A theoretically better trajectory is not useful if the controller takes too long to compute it. One part of the project therefore examines the tradeoff between horizon length, model fidelity, re-planning frequency, and the reaction time available before impact.

Validation plan

Controller development begins with MATLAB/Simulink and existing MuJoCo vehicle/contact studies. The architecture is intentionally structured so that real-time solver migration, higher-fidelity vehicle simulation, a future physical-damage metric, and eventual hardware validation can be added without rebuilding the rest of the system.

Related work

Alongside the vehicle-control research, I have worked on nonlinear state estimation, including CKF/EKF comparisons for localization and radar tracking and adaptive visual-inertial filtering for underwater navigation.

Research discussion

Interested in MPC, vehicle dynamics, or collision-aware control?