Mithun Babu1 Raghu Ram Theerthala1 Arun Kumar Singh2 Baladhurgesh B.P.1 Bharath Gopalakrishnan1 K. Madhava Krishna1 Shanti Medasani1
1 IIIT Hyderabad, India 2 TUIT, University of Tartu, Estonia 3 MathWorks, Inc.
Model Predictive control (MPC) is a powerful optimization strategy that can be efficiently employed in local reactive scenarios for autonomous driving. It is advantageous in providing a unified optimization framework that can system constraints during the design stage. However the key bottlenecks include requirement heavy computational resources and high fidelity in predictive nature of the algorithm. This work improves the computational performance of the Model predictive control in complex scenarios by employing Alternating Minimization(AM) framework. This framework splits the non-linear unicycle modle into convex-concave layers which results in faster minimization when compared to conventional joint optimization. The other contribution involves inclusion of Actuator dynamics into motion model by a data driven approach. The relation between actual body velocity and the commanded velocity is modeled as a first order approximation. This approximation provides an improvement in the vehicle behavior by implicitly compensating the actuator infidelities. This paper shows shows the improvement in the ego-vehicle behavior by testing in complicated scenarios like lane changing, occlusion during overtaking and sudden braking etc.,