Autonomous Nano-UAVs: An Extreme Edge Computing Case

Explore the cutting-edge research on autonomous nano-UAVs in this conference talk from tinyML EMEA 2022. Delve into the challenges and solutions for enabling extreme edge computing on miniature flying robots with stringent size and power constraints. Learn about innovative approaches using machine learning and deep neural networks to achieve onboard intelligence, including vision-based algorithms, quantization techniques, and data augmentation pipelines. Discover key projects like PULP-Dronet and PULP-Frontnet, addressing crucial questions about optimizing CNNs for autonomous navigation, improving generalization for human-robot interaction, and combining cyber-physical system state with vision-based CNNs. Gain insights into shrinking operation counts and memory footprints, enhancing tiny CNN performance, and integrating vision with state information. Examine thorough in-field evaluations and closed-loop robotic demonstrations that showcase the practical applications of these groundbreaking methodologies in nanorobotics.

Intro
Motivation/Vision
State-of-the-Art and Challenges
How to enable the extreme edge computing?
Full stack application: Autonomous navigation
On-board brain
PULP-Dronet: Shrinking and optimization
PULP-Dronet: Results
PULP-Dronet: In-field evaluation
PULP-Dronet evolution
PULP-Dronet v2: Results
PULP-Dronet v3: Tiny-PULP-Dronets
Full stack application: Human robot interaction
PULP-Frontnet: In-field evaluation
How to improve the generalization capability?
Background randomization: Pipeline
Background randomization: Testing setup
Background randomization: Results
Background randomization: In-field evaluation
How to improve the regression performance?
Vision-state fusion: In-field evaluation
Conclusion