Iterative Mechanical Design and Nonlinear Predictive Control for Autonomous Robotic Vehicles

Abstract

Small-scale autonomous ground vehicles operate in a regime where the conventional separa tion between mechanical design and control design is no longer tenable. The mass distribution of the chassis, the height of the perception assembly, and the electromagnetic environment of the on-board compute platform all enter the lateral dynamics model directly, and therefore enter the controller’s prediction model. A trajectory-tracking strategy formulated in isolation from these mechanical realities will fail in the field, and a chassis designed without reference to the controller’s stability margins will require ad-hoc retuning that erodes any performance claim. This thesis advances the position that the mechanical platform, the dynamics model, and the predictive controller must be co-designed, and demonstrates this co-design on the ODYSSIX autonomous robotic vehicle. The platform replaces the prevailing fixed-gain PID and statically toleranced mechanical paradigm with two coupled advances. Mechanically, the chassis is the product of a four iteration evolutionary process executed in CAD and Fused Deposition Modeling (FDM): a dual-polymer strategy– polylactic acid for dimensional precision in low-stress mounts, acrylonitrile butadiene styrene for impact resilience and passive vibrational damping in motor adjacent and mast-borne mounts– is paired with a uniform 60% gyroid infill that minimizes mass-at-height in accordance with the rotational inertia penalty I = mr2. Algorithmically, vehicle motion is described by an enhanced kinematic bicycle model in which the velocity state is grounded in direct AS5047P encoder measurement (replacing the non-physical aerodynamic drag term that customarily appears in such models) and the heading prediction incorporates a slip-angle correction to capture tire lateral deformation during cornering. This model is the prediction kernel of a Sequential Quadratic Programming (SQP)- Nonlinear Model Predictive Controller (NMPC) operating at 50Hz with a 20-step receding horizon. The controller’s principal contribution is an online adaptive cost-weight scheme: tracking error is decomposed into the Frenet frame of the reference path, and the lateral weight Qlat(v) is modulated as a function of vehicle speed, reflecting the v2 scaling of centripetal force demand. The result is a single, unified MPC formulation that adapts gracefully across the full operating speed envelope, eliminating the need for pre-compiled solver variants or regime-switching logic. The perception, embedded firmware, and Extended Kalman Filter (EKF) layers– developed by collaborators– supply the high-rate, electromagnetically clean state estimate that the controller requirestheir architecture is documented in this thesis to establish that environment, but is not claimed as the author’s contribution. Integration testing on curved reference corridors confirms stable closed-loop tracking, with the SQP-NMPC consistently solving in 3–8ms on the Jetson Orin Nano, well within the 20ms control-cycle budget. Orientation noise is sustained below 0.5◦ RMS following the deliberate excision of the magnetometer from the IMU’s fusion output, a decision motivated by an electromagnetic interference audit that is presented as part of the supporting case. The ODYSSIX platform is offered as evidence that mathematically verifiable, computationally tractable predictive control is realizable on emerging-market-grade autonomous hardware when, and only when, the mechanical and algorithmic layers are designed against one another rather than in series.