Training on the Fly: On-device Self-supervised Learning aboard Nano-drones within 20 mW

Miniaturized cyber-physical systems (CPSes) powered by tiny machine learning (TinyML), such as nano-drones, are becoming an increasingly attractive technology. Their small form factor (i.e., ~10cm diameter) ensures vast applicability, ranging from the exploration of narrow disaster scenarios to safe human-robot interaction. Simple electronics make these CPSes inexpensive, but strongly limit the computational, memory, and sensing resources available on board. In real-world applications, these limitations are further exacerbated by domain shift. This fundamental machine learning problem implies that model perception performance drops when moving from the training domain to a different deployment one. To cope with and mitigate this general problem, we present a novel on-device fine-tuning approach that relies only on the limited ultra-low power resources available aboard nano-drones. Then, to overcome the lack of ground-truth training labels aboard our CPS, we also employ a self-supervised method based on ego-motion consistency. Albeit our work builds on top of a specific real-world vision-based human pose estimation task, it is widely applicable for many embedded TinyML use cases. Our 512-image on-device training procedure is fully deployed aboard an ultra-low power GWT GAP9 System-on-Chip and requires only 1MB of memory while consuming as low as 19mW or running in just 510ms (at 38mW). Finally, we demonstrate the benefits of our on-device learning approach by field-testing our closed-loop CPS, showing a reduction in horizontal position error of up to 26% vs. a non-fine-tuned state-of-the-art baseline. In the most challenging never-seen-before environment, our on-device learning procedure makes the difference between succeeding or failing the mission.

Learning to Estimate the Pose of a Peer Robot in a Camera Image by Predicting the States of its LEDs

We consider the problem of training a fully convolutional network to estimate the relative 6D pose of a robot given a camera image, when the robot is equipped with independent controllable LEDs placed in different parts of its body. The training data is composed by few (or zero) images labeled with a ground truth relative pose and many images labeled only with the true state (\textsc{on} or \textsc{off}) of each of the peer LEDs. The former data is expensive to acquire, requiring external infrastructure for tracking the two robots; the latter is cheap as it can be acquired by two unsupervised robots moving randomly and toggling their LEDs while sharing the true LED states via radio. Training with the latter dataset on estimating the LEDs' state of the peer robot (\emph{pretext task}) promotes learning the relative localization task (\emph{end task}). Experiments on real-world data acquired by two autonomous wheeled robots show that a model trained only on the pretext task successfully learns to localize a peer robot on the image plane; fine-tuning such model on the end task with few labeled images yields statistically significant improvements in 6D relative pose estimation with respect to baselines that do not use pretext-task pre-training, and alternative approaches. Estimating the state of multiple independent LEDs promotes learning to estimate relative heading. The approach works even when a large fraction of training images do not include the peer robot and generalizes well to unseen environments.

Predicting the Intention to Interact with a Service Robot:the Role of Gaze Cues

For a service robot, it is crucial to perceive as early as possible that an approaching person intends to interact: in this case, it can proactively enact friendly behaviors that lead to an improved user experience. We solve this perception task with a sequence-to-sequence classifier of a potential user intention to interact, which can be trained in a self-supervised way. Our main contribution is a study of the benefit of features representing the person's gaze in this context. Extensive experiments on a novel dataset show that the inclusion of gaze cues significantly improves the classifier performance (AUROC increases from 84.5% to 91.2%); the distance at which an accurate classification can be achieved improves from 2.4 m to 3.2 m. We also quantify the system's ability to adapt to new environments without external supervision. Qualitative experiments show practical applications with a waiter robot.

Resource-Aware Collaborative Monte Carlo Localization with Distribution Compression

Global localization is essential in enabling robot autonomy, and collaborative localization is key for multi-robot systems. In this paper, we address the task of collaborative global localization under computational and communication constraints. We propose a method which reduces the amount of information exchanged and the computational cost. We also analyze, implement and open-source seminal approaches, which we believe to be a valuable contribution to the community. We exploit techniques for distribution compression in near-linear time, with error guarantees. We evaluate our approach and the implemented baselines on multiple challenging scenarios, simulated and real-world. Our approach can run online on an onboard computer. We release an open-source C++/ROS2 implementation of our approach, as well as the baselines

On-device Self-supervised Learning of Visual Perception Tasks aboard Hardware-limited Nano-quadrotors

Sub-\SI{50}{\gram} nano-drones are gaining momentum in both academia and industry. Their most compelling applications rely on onboard deep learning models for perception despite severe hardware constraints (\ie sub-\SI{100}{\milli\watt} processor). When deployed in unknown environments not represented in the training data, these models often underperform due to domain shift. To cope with this fundamental problem, we propose, for the first time, on-device learning aboard nano-drones, where the first part of the in-field mission is dedicated to self-supervised fine-tuning of a pre-trained convolutional neural network (CNN). Leveraging a real-world vision-based regression task, we thoroughly explore performance-cost trade-offs of the fine-tuning phase along three axes: \textit{i}) dataset size (more data increases the regression performance but requires more memory and longer computation); \textit{ii}) methodologies (\eg fine-tuning all model parameters vs. only a subset); and \textit{iii}) self-supervision strategy. Our approach demonstrates an improvement in mean absolute error up to 30\% compared to the pre-trained baseline, requiring only \SI{22}{\second} fine-tuning on an ultra-low-power GWT GAP9 System-on-Chip. Addressing the domain shift problem via on-device learning aboard nano-drones not only marks a novel result for hardware-limited robots but lays the ground for more general advancements for the entire robotics community.

High-throughput Visual Nano-drone to Nano-drone Relative Localization using Onboard Fully Convolutional Networks

Relative drone-to-drone localization is a fundamental building block for any swarm operations. We address this task in the context of miniaturized nano-drones, i.e., 10cm in diameter, which show an ever-growing interest due to novel use cases enabled by their reduced form factor. The price for their versatility comes with limited onboard resources, i.e., sensors, processing units, and memory, which limits the complexity of the onboard algorithms. A traditional solution to overcome these limitations is represented by lightweight deep learning models directly deployed aboard nano-drones. This work tackles the challenging relative pose estimation between nano-drones using only a gray-scale low-resolution camera and an ultra-low-power System-on-Chip (SoC) hosted onboard. We present a vertically integrated system based on a novel vision-based fully convolutional neural network (FCNN), which runs at 39Hz within 101mW onboard a Crazyflie nano-drone extended with the GWT GAP8 SoC. We compare our FCNN against three State-of-the-Art (SoA) systems. Considering the best-performing SoA approach, our model results in an R-squared improvement from 32 to 47% on the horizontal image coordinate and from 18 to 55% on the vertical image coordinate, on a real-world dataset of 30k images. Finally, our in-field tests show a reduction of the average tracking error of 37% compared to a previous SoA work and an endurance performance up to the entire battery lifetime of 4 minutes.

Self-Supervised Learning of Visual Robot Localization Using LED State Prediction as a Pretext Task

We propose a novel self-supervised approach for learning to visually localize robots equipped with controllable LEDs. We rely on a few training samples labeled with position ground truth and many training samples in which only the LED state is known, whose collection is cheap. We show that using LED state prediction as a pretext task significantly helps to learn the visual localization end task. The resulting model does not require knowledge of LED states during inference. We instantiate the approach to visual relative localization of nano-quadrotors: experimental results show that using our pretext task significantly improves localization accuracy (from 68.3% to 76.2%) and outperforms alternative strategies, such as a supervised baseline, model pre-training, and an autoencoding pretext task. We deploy our model aboard a 27-g Crazyflie nano-drone, running at 21 fps, in a position-tracking task of a peer nano-drone. Our approach, relying on position labels for only 300 images, yields a mean tracking error of 4.2 cm versus 11.9 cm of a supervised baseline model trained without our pretext task. Videos and code of the proposed approach are available at https://github.com/idsia-robotics/leds-as-pretext

A Sim-to-Real Deep Learning-based Framework for Autonomous Nano-drone Racing

Autonomous drone racing competitions are a proxy to improve unmanned aerial vehicles' perception, planning, and control skills. The recent emergence of autonomous nano-sized drone racing imposes new challenges, as their ~10cm form factor heavily restricts the resources available onboard, including memory, computation, and sensors. This paper describes the methodology and technical implementation of the system winning the first autonomous nano-drone racing international competition: the IMAV 2022 Nanocopter AI Challenge. We developed a fully onboard deep learning approach for visual navigation trained only on simulation images to achieve this goal. Our approach includes a convolutional neural network for obstacle avoidance, a sim-to-real dataset collection procedure, and a navigation policy that we selected, characterized, and adapted through simulation and actual in-field experiments. Our system ranked 1st among seven competing teams at the competition. In our best attempt, we scored 115m of traveled distance in the allotted 5-minute flight, never crashing while dodging static and dynamic obstacles. Sharing our knowledge with the research community, we aim to provide a solid groundwork to foster future development in this field.

Self-Supervised Prediction of the Intention to Interact with a Service Robot

A service robot can provide a smoother interaction experience if it has the ability to proactively detect whether a nearby user intends to interact, in order to adapt its behavior e.g. by explicitly showing that it is available to provide a service. In this work, we propose a learning-based approach to predict the probability that a human user will interact with a robot before the interaction actually begins; the approach is self-supervised because after each encounter with a human, the robot can automatically label it depending on whether it resulted in an interaction or not. We explore different classification approaches, using different sets of features considering the pose and the motion of the user. We validate and deploy the approach in three scenarios. The first collects $3442$ natural sequences (both interacting and non-interacting) representing employees in an office break area: a real-world, challenging setting, where we consider a coffee machine in place of a service robot. The other two scenarios represent researchers interacting with service robots ($200$ and $72$ sequences, respectively). Results show that, even in challenging real-world settings, our approach can learn without external supervision, and can achieve accurate classification (i.e. AUROC greater than $0.9$) of the user's intention to interact with an advance of more than $3$s before the interaction actually occurs.

Sim-to-Real Vision-depth Fusion CNNs for Robust Pose Estimation Aboard Autonomous Nano-quadcopter

Nano-quadcopters are versatile platforms attracting the interest of both academia and industry. Their tiny form factor, i.e., $\,$10 cm diameter, makes them particularly useful in narrow scenarios and harmless in human proximity. However, these advantages come at the price of ultra-constrained onboard computational and sensorial resources for autonomous operations. This work addresses the task of estimating human pose aboard nano-drones by fusing depth and images in a novel CNN exclusively trained in simulation yet capable of robust predictions in the real world. We extend a commercial off-the-shelf (COTS) Crazyflie nano-drone -- equipped with a 320$\times$240 px camera and an ultra-low-power System-on-Chip -- with a novel multi-zone (8$\times$8) depth sensor. We design and compare different deep-learning models that fuse depth and image inputs. Our models are trained exclusively on simulated data for both inputs, and transfer well to the real world: field testing shows an improvement of 58% and 51% of our depth+camera system w.r.t. a camera-only State-of-the-Art baseline on the horizontal and angular mean pose errors, respectively. Our prototype is based on COTS components, which facilitates reproducibility and adoption of this novel class of systems.