← BACK TO PROJECTS
Spatial Hyper-MPC with Neural Policy Distillation
#Autonomous Controls
#Model Predictive Control
#Neural Policy Distillation
#OSQP Solver
#Joint EKF
#6-DOF Vehicle Dynamics
Engineered an ultra-fast, high-dimensional control framework for active vehicle suspension systems subjected to aggressive spatial road disturbances.
Technical Architecture & Innovations
- Formulation of 6-DOF Dynamic Suspension System: Modeled full vehicle vertical, pitch, and roll dynamics coupled with non-linear actuator constraints and road profile kinematics.
- Hybrid Neural-MPC Controller:
- Implemented an online Model Predictive Controller formulated as a convex Quadratic Program (QP) solved via the
OSQPsolver. - Coupled an online Joint Extended Kalman Filter (EKF) to estimate unmeasured chassis states and dynamic tire normal forces in real time.
- Implemented an online Model Predictive Controller formulated as a convex Quadratic Program (QP) solved via the
- Neural Policy Distillation for Real-Time Inference:
- Distilled a computationally demanding 40-step horizon MPC into a compact 4,736-parameter deep neural policy network trained over 20,000 expert simulation trajectories using behavioral cloning and DAgger.
- Reduced solver latency from 15ms (unsuitable for high-frequency chassis stabilization) to under 0.3ms, enabling microsecond-level execution.
- Rigorous Benchmarking: Validated against extensively tuned Multi-Input Multi-Output (MIMO) PID baselines across 3 standardized terrain profiles, demonstrating up to 3× reduction in peak angular velocities (pitch and roll).