Problem statement
In this Master Thesis in Software Engineering, the focus will be on defining and validating a benchmarking framework for AI models deployed on micro-drones. Due to the constrained nature of the micro-drone platforms, it offers challenges regarding limited processing power, memory, energy, payload capacity, and flight time. These limitations make it difficult to determine whether an AI or machine-learning model that performs well in simulation will operate efficiently and reliably also on the target hardware.
The central research question is: which benchmarking metrics are the most effective for measuring the performance of AI models deployed on resource-constrained micro-drones? The work should identify suitable key performance indicators (KPIs), with a need to understand and investigate the trade-offs between model performance and drone constraints, to finally, define a pragmatic test scenario that fits the available timebox and hardware availability.
The project will build on simulation environment, simulation data and will be tested via existing hardware and software environment, with provided drone components, programming and control interfaces, Machine Learning/AI drone knowledge, and drone related simulation assets. The evaluation will therefore connect model-level measurements with the realities of embedded deployment and bounded flight testing.
Proposed solution
The proposed solution is to design and implement a repeatable benchmarking framework for testing AI models into a micro-drone target, first with simulation, but also with deployment of the model on a real micro-drone platform, this test framework will capture relevant measurements, and will support a structured comparison of performance indicators.
The work will include: Benchmark definition: Identify and define metrics which characterize the model quality, execution performance, resource usage, energy-related trade-offs and operational behavior.
- Deployment pipeline: Review and adapt the current pipeline for integrating and running selected ML inference models on micro-drone hardware targets.
- Model and scenario selection: Select suitable available models and define one controlled, fingerprint scenario and a test environment for meaningful and reproducible evaluation.
- Framework implementation: Develop the measurement, logging, analysis, and reporting workflow needed to extract benchmarking parameters from the micro-drone.
- Validation: Run a focused test suite, first on simulation, then, compare results across relevant configurations, and assess how the chosen metrics support practical deployment decisions.
The expected result is a Proof-of-Concept benchmarking framework and a documented evaluation method for micro-drone platforms. The work is deliberately bounded; it does not aim to create an all-case autonomous flight solution but to create the bases within a specific micro-drone flight scenario for benchmark testing. The student will have the opportunity to shape the final selection of models, metrics, and test scenarios based on feasibility and findings during the project.
Your profile
In order to be successful in this project, we think you are:
- A student in Information Technology, Computer Science, Electronics, Mathematics or Physics.
- Have relevant coursework or knowledge in data science, control systems, and AI.
- Experienced with or have at least basic programming knowledge in Python, C++ or similar languages.
- Self-driven, able to take initiative and move the project forward independently.
A person with team spirit, social skills, and genuine curiosity for robotics, AI and autonomous systems.
Two students completing 30 credits each (20 weeks) onsite at the Lund office
Please note: Applications from students located in the areas of Lund and Malmö are prefered.
Contact: Diego Medina, mail [email protected]