Control Algorithms and Hardware Accelerations for Unmanned Aerial Vehicles: Implementation and Numerical Precision Analysis for MPPI Acceleration

Control Algorithms and Hardware Accelerations for Unmanned Aerial Vehicles: Implementation and Numerical Precision Analysis for MPPI Acceleration This thesis describes the development and optimization of an advanced drone control system based on the Model Predictive Path Integral (MPPI) approach, with a focus on analyzing numerical precision for real-time applications on accelerated hardware. The work was conducted at the INRIA research center in France, in collaboration with the Rainbow team, which specializes in robotics, and the Taran team, focused on hardware research. Real-time control of autonomous systems represents one of the main challenges in modern robotics, as it requires both high computational efficiency and precision. The primary objective of this work was to implement an MPPI controller on a quadrotor drone equipped with an NVIDIA Jetson Orin platform, leveraging GPU acceleration via CUDA and the JAX framework, to evaluate the impact of numerical precision on computational performance and control quality. At the same time, the need to run these algorithms on resource-constrained hardware makes the choice of numerical representation a crucial aspect of the project. The main contribution of this thesis consists of the design and implementation of a complete MPPI controller, integrated into the TeleKyb3 framework. The integration required a careful analysis of the Jetson platform’s architectural specifications and the characteristics of the JAX framework to ensure compatibility, efficiency, and real-time execution. A significant portion of the work was dedicated to the systematic analysis of different numerical precisions (float16 and float32) through tests conducted in a simulation environment. By comparing computation time and tracking error in different scenarios, it was possible to obtain useful insights for more informed hardware design decisions. Experimental results show that the use of float16 significantly reduces computation time compared to float32, while ensuring essentially equivalent control performance. In terms of tracking error, both precisions offer comparable results, without revealing a clear superiority of one over the other. This conclusion is particularly relevant for future FPGA implementations, where the use of float16 arithmetic could enable more compact and efficient hardware architectures. A float16-based FPGA implementation could operate at higher frequencies than a similar float32 implementation, ensuring not only reduced processing times but also a higher control update rate. In summary, this work proposes a validated approach that combines controller implementation, numerical precision analysis, and performance evaluation into a single integrated framework both in simulation and the real world. The results obtained represent a concrete and innovative step towards more efficient MPPI controllers on dedicated hardware, demonstrating the feasibility and benefits of using reduced numerical precision for real-time control applications in robotics.