Automated Tuning of End-to-end Neural Flight Controllers for Autonomous Nano-drones

Tiny-PULP-Dronets: Squeezing Neural Networks for Faster and Lighter Inference on Multi-Tasking Autonomous Nano-Drones

“…We use the RocketLogger data logger [13] (64 ksps) to separately plot the power waveforms of the GAP8's main core, the Fabric Controller (FC), and its 8-core parallel cluster (CL). We test two SoC configurations: the so called energy-efficient configuration [10], whose operating point is FC@50 MHz, CL@100 MHz, and VDD@1 V, and the maximum performance settings according to the GAP8's datasheet, which is FC@250 MHz, CL@175 MHz, VDD@1.2 V.…”

“…To date, nano-drones can accomplish individual intelligence autonomous navigation tasks by leveraging deep learningbased (DL) algorithms -spanning a broad spectrum of complexity and sensorial input [2], [8]- [10]. In [8], Kooi and Babuška use a deep reinforcement learning approach with proximal policy optimization for the autonomous landing of a nano-drone on an inclined surface.…”

“…An example of these newgeneration devices is given by the 9-core GWT GAP8 1 , a parallel-ultra-low-power (PULP) processor. This processor was previously exploited on UAVs in the PULP-Dronet project [10], leading to a fully-programmable end-to-end visual-based autonomous navigation engine for nano-drones. PULP-Dronet is a single CNN capable of navigating a 27grams nano-drone in both indoor and outdoor environments by predicting a collision probability, for obstacle avoidance, and a steering angle, to keep the drone within a lane.…”

“…Ultimately, the Tiny-PULP-Dronets, with a minimum model size of only 6.4 kB and a maximum throughput of more than 160 frame/s, enable higher reactivity on the autonomous navigation task, and leave sufficient memory and compute headroom for onboard multitasking intelligence even at the nano-sized scale. [10] is a ResNet-based [11] CNN made of three residual blocks (ResBlocks), where each one consists of a main branch, performing two 3 × 3 convolutions, and a parallel by-pass (Byp), performing one 1×1 convolution. This CNN produces a steering angle (regression) and a collision probability (classification) output.…”