Exploring a Learning-Based Approach to Motion Planning and Control for Over-Actuated Autonomous Vehicles

Paper in proceeding, 2026

The motion planning and control of over-actuated autonomous
electric vehicles (AEVs) relies on coordinating
redundant actuators, such as individual-wheel torque vectoring,
through the weighting of the controller cost function
terms. This weighting is commonly tuned and fixed
for a single situation, although the most effective actuator
distribution changes with road curvature and with the uneven
per-wheel loads induced by load transfer in corners.
This paper combines reinforcement learning (RL) with
model predictive control (MPC) by adapting the control
weightings online. An RL policy is trained to map the instantaneous
vehicle and load state and the upcoming road
curvature to a load-aware control weighting matrix, which
is supplied to the MPC at every control step while the vehicle
dynamics and the safety constraints remain inside
the controller. The method is evaluated in closed-loop
simulations on a handling track against both a manually
tuned default-weighting MPC baseline and an optimally
tuned one. In both cases the RLMPCconsistently reduces
the lap time while keeping the vehicle on the track, and
the gain persists even against the optimally tuned baseline.
Under parameter mismatch, RL MPC recovers an
on-track trajectory where the MPC controller leaves the
track, reducing the peak lateral deviation from 2.83m to
2.17m and the lap time by 0.54 s.