Context & background:

Robots are physical agents that interact with their physical environment. Accordingly, their sensorimotor capabilities are essential and largely determine the activities that robots can perform. In recent years, great progress has been made in sensory capabilities thanks to significant advances in machine learning and dedicated hardware. In contrast, much less progress has been made in motor skills. Examples of promising approaches in the current scientific literature are Model Predictive Control (MPC) [1] and Model Predictive Path Integral (MPPI) control [2], where control actions are optimized over a finite time horizon, considering the time evolution of robot dynamics to optimize a given cost or reward function that describes the robot motion. Such algorithms are particularly suited for optimizing control trajectories and planning horizons in real time due to their ability to handle dynamic environments.

From a control perspective, planning a horizon that is as long as possible to manage complex trajectories while considering the environment is essential. Additionally, maintaining a high control frequency is crucial to meet the real-time demands imposed by real-world physics and, if necessary, to adjust the sequence of movements. In the resource-constrained context of small-scale UAVs, such control algorithms are crucial as they enable optimal trajectory generation and real-time decision-making in complex, dynamic, and uncertain environments. However, particularly for battery-powered UAVs, achieving a high control frequency while planning for a long horizon is difficult due to limited computational power and energy constraints [3], and conventional GPU acceleration often requires excessive energy consumption.

In recent years, hardware acceleration [4] has become increasingly popular, using dedicated platforms such as FPGAs (Field Programmable Gate Arrays) and ASICs (Application-specific Integrated Circuits), increasing energy efficiency by orders of magnitude [5]. However, dedicated hardware acceleration for small-scale UAV control has not been proposed.

The Phd is in collaboration between the computer architecture team (TARAN) and the robotics team (RAINBOW) at Inria Centre at Rennes University. 

Prospective candidates must manifest their interest before June 1, 2025, to prepare for the doctoral audition, scheduled on June 12, 2025 (remotely or in person).

Contact people:
Marcello Traiola, marcello.traiola@inria.fr
Marco Tognon, marco.tognon@inria.fr
Tommaso Belvedere, tommaso.belvedere@inria.fr

[1] E. F. Camacho and C. Bordons, Model Predictive control. in Advanced Textbooks in Control and Signal Processing. London: Springer, 2007. doi: 10.1007/978-0-85729-398-5.
[2] G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Aggressive driving with model predictive path integral control,” in 2016 IEEE International Conference on Robotics and Automation (ICRA), May 2016, pp. 1433–1440. doi: 10.1109/ICRA.2016.7487277.
[3] K. Nguyen, S. Schoedel, A. Alavilli, B. Plancher, and Z. Manchester, “TinyMPC: Model-Predictive Control on Resource-Constrained Microcontrollers,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), May 2024, pp. 1–7. doi: 10.1109/ICRA57147.2024.10610987.
[4] W. J. Dally, Y. Turakhia, and S. Han, “Domain-specific hardware accelerators,” Commun ACM, vol. 63, no. 7, pp. 48–57, Jun. 2020, doi: 10.1145/3361682
[5] J. L. Hennessy and D. A. Patterson, “A new golden age for computer architecture,” Commun ACM, vol. 62, no. 2, pp. 48–60, Jan. 2019, doi: 10.1145/3282307.

This Ph.D. thesis aims to use algorithm-specific custom hardware acceleration to implement efficient real-time control for UAVs with long prediction horizons and high control frequencies. The structure of the control algorithms is complex and sensitive to numerical errors or reduced arithmetic precision. Thus, applying a hardware-algorithm Co-design approach is necessary, i.e., adapting the control algorithms to the hardware and designing the hardware to suit the control algorithms optimally. 

After a detailed study of UAV state-of-the-art control algorithms, the student will identify HW acceleration opportunities, such as parallelization, pipelining, and data specialization. The student will apply co-design approaches to realize efficient accelerators, utilizing the control algorithms’ properties to improve the hardware while adjusting the algorithms to the hardware’s characteristics. Simulations will be carried out to validate the proposed approaches and prepare the final integration in the UAV platform, which is already available to the RAINBOW team.