The central question the Delft team set out to answer was deceptively simple: how much performance can be extracted from a robot with deliberately minimal sensing, computing and energy resources? The question matters because drone racing sits at the extreme end of what autonomous systems can do. At racing speeds, the margin for error shrinks to almost nothing. A delay of even a few tens of milliseconds in perceiving a gate or issuing a control command can mean clipping a gate, colliding with an obstacle, or losing stability entirely and crashing. Any system that can fly reliably under those constraints has solved a problem that transfers directly to other fast-moving robots, from self-driving vehicles to mobile and humanoid platforms.

MonoRace’s most striking feature is what it does without. Previous leading autonomous racing systems have relied on richer sensor suites, often including GPS, external motion-capture tracking or multiple cameras, to pin down the drone’s position with high precision. The Delft system, developed at the university’s Micro Air Vehicles Laboratory in the Faculty of Aerospace Engineering, uses only one forward-facing camera and an inertial measurement unit, or IMU, for everything: perception, navigation and control. There is no GPS and no external tracking. The drone must figure out where it is and where it is going entirely from what it sees and how it feels its own motion, in real time, while hurtling through a course designed by other people.

Seeing the track through a single lens at racing speed is a formidable perception problem. The images arriving at the camera are blurry and deformed by the drone’s rapid motion, and the neural network has only a fraction of a second to identify the racing gates and estimate where the drone sits relative to them. MonoRace handles this by splitting the work between two complementary sources of information. A neural network examines the camera images and detects the gates, while the onboard IMU, combined with a model of the drone’s own dynamics, estimates the vehicle’s position and motion between visual observations. By fusing these streams, the system maintains a reliable picture of its state on the track even when the camera alone would be insufficient. Stavrow Bahnam, the study’s main author, described the challenge plainly: flying at very high speed with only one camera is difficult because the images are blurry and deformed and the AI has very little time to determine precisely where the drone is. The team’s solution, he said, was a robust way of combining what the camera sees with the drone’s own motion information, allowing it to navigate the track at high speed without GPS. The system also had to tolerate the imperfections of real flight, including camera interference and sensor saturation, conditions that rarely appear in simulation but are routine in the air.

The second major innovation lies in how the drone is controlled. Conventional flight stacks are layered: a planner proposes a trajectory, intermediate controllers translate that trajectory into attitude commands, and lower-level loops convert those commands into motor signals. Each layer adds latency and each abstraction discards some information about what the platform can actually do. MonoRace discards the layers. A small neural network, trained with reinforcement learning, takes raw state information and directly generates the commands sent to the drone’s motors. The network runs onboard the flight controller, the tiny computing chip at the heart of the aircraft, and issues control commands 500 times per second. Because it was trained through reinforcement learning, the network internalised the drone’s physical limitations during training, learning to exploit the full capabilities of the airframe without exceeding them. Robin Ferede, an author of the study, explained the rationale: instead of first planning a desired flight path and conventional commands, the neural network learns to directly control the motors, which allows the controller to react extremely quickly and exploit the drone’s capabilities while remaining small enough to run onboard.

That combination of speed and compactness is what makes the result scientifically significant rather than merely spectacular. Reinforcement learning has produced impressive controllers before, but many of them depend on heavyweight networks, powerful onboard computers or offboard infrastructure that cannot be carried by small, energy-constrained robots. MonoRace demonstrates that a network small enough to fit on a standard flight controller, running at 500 hertz, can execute champion-level racing. The efficiency is not an aesthetic choice; it is the enabling condition. A racing drone has strict limits on weight and power, and every gram of computer or battery spent on processing is a gram not spent on performance. The Delft team treated those constraints as the design problem itself rather than as obstacles to be engineered away.

The validation environment was chosen to make the test as unforgiving as possible. The 2025 Abu Dhabi Autonomous Drone Racing Competition was organised and operated externally: the drone hardware and the racing environment were designed and run by the competition organisers, not by the Delft researchers. That meant the system had to perform on equipment and a track the research team did not control, with no opportunity to tune the perception pipeline to familiar gate colours, lighting conditions or course geometry. This is a crucial distinction from laboratory demonstrations, where researchers typically race on courses they built themselves and on hardware they know intimately. Despite those constraints, the TU Delft drone reached speeds of up to 100 kilometres per hour and finished ahead of all other autonomous teams and three former world champion human pilots. Christophe De Wagter, one of the researchers, framed the philosophy behind the effort: rather than giving the robot increasingly powerful sensors and computers, the team wanted to find out how much intelligence could be extracted from very limited resources. Drone racing, he said, provided an extreme environment in which to test these ideas outside the laboratory, pushing the group to develop AI that was not only fast but also robust and able to adapt when external reference measurements were unavailable.

Beating human champions is more than a symbolic victory. Elite human drone pilots fly by integrating visual cues, learned track knowledge and finely tuned reflexes, and at racing speeds their reaction times and anticipatory skills represent a genuine benchmark of dynamic control. An autonomous system that surpasses them while carrying less sensing than a human’s own two eyes suggests that learned neural-network controllers, when trained well and run fast enough, can outperform biological perception-action loops in tasks at the edge of physical feasibility. It also confirms a trend visible across robotics: end-to-end learned control, once dismissed as too brittle for real hardware, can now hold its own in competitive, physical, high-stakes settings.

For Guido de Croon, professor at Delft University of Technology, the significance of the work extends well beyond making drones faster. Autonomous racing, in his view, is a way of pushing robotic AI to its limits. In the current study, the team made an AI efficient enough to control a drone travelling at 100 kilometres per hour using a single camera and a small onboard processor, and the same AI could be applied to many different real-world robotic systems. The optimisation objective, he noted, is flexible: here the researchers optimised for speed, but the same approach could be used to prioritise efficiency or safety, for example by performing smooth, energy-optimal control on systems ranging from small vacuum-cleaning robots to huge ship-to-shore cranes. The point is that the underlying control principles are not specific to racing; they are about making fast, reliable decisions with limited onboard resources, which is the defining constraint of nearly all autonomous robotics.

The broader implications reach across the field. Autonomous road vehicles, delivery drones, inspection robots and future humanoid systems all face versions of the same problem MonoRace solved: perceive the environment with limited sensors, estimate one’s own state without external references, and act quickly enough to matter, all within a tight computing and energy budget. A methodology that extracts champion-level performance from a single camera and a millimetre-scale processor offers a template for doing exactly that. The Delft result suggests that the future of capable autonomous robots may depend less on ever-larger sensor arrays and computers, and more on the intelligence that can be distilled into small, fast, robust networks trained to master the physics of the machines they control. In drone racing, that idea has now been proven at 100 kilometres per hour, in front of the world’s best human pilots, on a course its creators never built.

Subject of Research: Neural-network-based autonomous drone racing control using a single camera

Article Title: The science behind the world’s fastest autonomous racing drone

Article References: The science behind the world’s fastest autonomous racing drone. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: autonomous drones, drone racing, MonoRace, neural networks, reinforcement learning, computer vision, TU Delft, Science Robotics, robotics, embedded AI, monocular camera, flight control

Millimetre-wave radar, the same class of technology used in automotive collision avoidance and next-generation wireless communications, emits radio waves in the millimetre band and listens for reflections. When those reflections bounce off a moving person, the radar’s signal processing pipeline produces a point cloud: a scattered set of three-dimensional coordinates tracing the surfaces and limbs that returned the signal. Unlike the dense, well-structured point clouds produced by lidar or depth cameras, radar point clouds are notoriously sparse, often containing only a handful of points per frame, and they are contaminated by multipath reflections and hardware noise. That sparsity and irregularity is precisely what has made them so difficult for deep learning systems to interpret.

The Aston team’s insight begins with a widely used building block of point-cloud deep learning called EdgeConv, introduced originally in the Dynamic Graph CNN architecture. EdgeConv treats each point as a node in a graph, connects it to its k nearest neighbours, and computes features along the edges connecting them, capturing the local geometry of the scene. The problem, as the researchers point out, is that conventional EdgeConv is deterministic and indiscriminate: it always uses every one of its k nearest neighbours in every forward pass, faithfully incorporating noisy, redundant or uninformative points into the representation. On clean, dense benchmark datasets from 3D vision, that determinism is a virtue. On sparse, noisy radar data, it becomes a liability, because the same corrupted neighbourhoods poison the features pass after pass.

SEdgeNet’s central contribution is a redesigned operator the authors call Stochastic Edge Convolution, or SEdgeConv, which is built from two complementary mechanisms. The first, Stochastic Neighbour Sampling, randomly selects only a subset of each point’s k nearest neighbours during each forward pass, governed by a sampling ratio that determines what fraction of the neighbourhood is used. The idea echoes dropout-style regularisation techniques from the graph neural network literature, such as DropEdge: by forcing the network to produce useful representations from constantly varying, randomly drawn subsets of neighbours, the model is discouraged from leaning on any single noisy point or fixed local configuration. Crucially, the researchers note, this is not a permanent loss of information. Across the many forward passes of training, the full neighbourhood structure remains accessible; the stochasticity simply prevents overreliance on any particular slice of it.

The second mechanism, Edge-wise Factorised Convolution, tackles the computational side of the problem. Standard EdgeConv applies a dense transformation across all edge feature channels jointly, which is expensive. EFC instead adapts the depthwise separable convolution paradigm, originally developed to make convolutional networks efficient on regular image grids, to the irregular edge features produced by stochastically sampled neighbourhoods. Each channel of the edge feature tensor is first transformed independently, and the per-channel responses are then fused through a pointwise linear projection. The authors stress that the efficiency gain comes from coupling this factorised processing with the dynamically varying graph structure created by the stochastic sampling, rather than from the factorisation scheme alone. Neighbour features are then aggregated with symmetric max pooling, yielding a permutation-invariant representation that does not depend on the ordering of points.

Around these operators, SEdgeNet stacks multiple SEdgeConv layers and concatenates their outputs along the channel dimension, forming a multi-scale embedding that preserves both fine-grained local geometry and more abstract global structure. A dual pooling strategy, combining adaptive max pooling with adaptive average pooling, distils this embedding into a fixed-length global descriptor; max pooling emphasises salient structures while average pooling encodes overall shape statistics. The pooled vectors feed a three-layer multilayer perceptron classifier that outputs the predicted activity class. The model is trained with stochastic gradient descent, a cosine annealing learning rate schedule, cross-entropy loss and early stopping, with all reported results averaged over three independent training runs using different random seeds.

The team evaluated SEdgeNet on two public mmWave radar datasets. The first, MMActivity, contains 93 minutes of point cloud data covering five activities performed by two subjects, captured with a Texas Instruments IWR1443BOOST sensor and processed into 12,097 training and 3,538 test samples using two-second windows. The second, MiliPoint, is far more demanding: 49 activities performed by 11 subjects, recorded with a TI IWR1843 radar and split 80-10-10 for training, validation and testing. On MMActivity, SEdgeNet outperformed the best previous model by 0.72 percent in accuracy, 0.71 percent in precision, 2.7 percent in recall and 0.71 percent in F1 score. On MiliPoint, the gains were larger still: 5.42 percent in accuracy, 4.91 percent in precision, 4.81 percent in recall and 5.26 percent in F1 score, establishing new state-of-the-art results on both benchmarks.

What makes those numbers remarkable is that they were achieved while using fewer computational resources than every baseline compared. SEdgeNet has the fewest parameters and the lowest multiply-accumulate count among the tested methods. Against DGCNN, the most efficient baseline, it uses 0.12 million fewer parameters and 0.02 giga-MACs less computation while achieving dramatically higher accuracy. Against AGConv, the most accurate baseline, it improves accuracy by 5.42 percent while cutting parameters by 0.52 million and computation by 8.43 giga-MACs. In controlled experiments swapping SEdgeConv for standard EdgeConv within an otherwise identical architecture, the stochastic version improved accuracy by 9.4 percent on MiliPoint and 0.44 percent on MMActivity, and required 3.08 giga-MACs less computation at a neighbourhood size of five. Ablation studies confirmed that both SNS and EFC contribute independently, and that combining them yields the best accuracy on both datasets alongside the lowest computational cost of 0.34 giga-MACs.

The hyperparameter analysis revealed instructive differences between the two datasets. MMActivity, with only five relatively separable classes, peaked with two layers and 35 nearest neighbours, reaching 98.12 percent accuracy, and its performance was dominated by network depth rather than neighbourhood size. MiliPoint, with 49 harder-to-distinguish classes, required a larger neighbourhood of 50 neighbours and five layers to reach 85.70 percent, suggesting that harder classification tasks demand both wider receptive fields and deeper networks to extract sufficiently discriminative geometric structure. The optimal sampling ratio also differed, at 0.6 for MMActivity and 0.3 for MiliPoint, though the authors found that a range of 0.3 to 0.5 delivers near-optimal accuracy on both datasets and recommend it as a practical default. Because the stochastic sampling remains active at inference time, the team also examined a Monte Carlo-style deployment strategy: averaging predictions over multiple forward passes. On MMActivity, accuracy rose from 97.72 percent with a single pass to a peak of 98.34 percent at six passes, at a still-modest cost of 2.04 giga-MACs.

Perhaps most compelling is the demonstration that the system runs in real time on embedded hardware. The researchers deployed their trained pipeline on a Jetson Nano mounted on a Robotino robot, with an AWR1843BOOST radar feeding data over USB. In a proof-of-concept trial, a single subject performed sitting, standing and walking freely in an indoor room, and the system achieved 99 percent classification accuracy, recognising actions under both lit and dark conditions. Average preprocessing latency was 12.18 milliseconds and model inference averaged 15.14 milliseconds, for a combined end-to-end latency of 27.32 milliseconds, comfortably within real-time requirements. The authors are careful to frame this as a feasibility demonstration rather than a comprehensive deployment study, noting that broader evaluation across multiple subjects, more activities and varying distances remains future work. Even so, the combination of privacy preservation, darkness-robust sensing, state-of-the-art accuracy and embedded real-time performance positions SEdgeNet as a promising foundation for assistive robotics, ambient intelligence and independent-living support, where the goal is to watch over people’s wellbeing without ever truly watching them.

Subject of Research: Graph neural network for human activity recognition from sparse millimetre-wave radar point clouds

Article Title: SEdgeNet: stochastic edge network for human activity recognition using sparse point cloud

Article References: Zakka, V. G., Manso, L. J., & Dai, Z. (2026). SEdgeNet: stochastic edge network for human activity recognition using sparse point cloud. Journal of Ambient Intelligence and Humanized Computing. https://doi.org/10.1007/s12652-026-05133-z

Image Credits: AI Generated

DOI: 10.1007/s12652-026-05133-z

Keywords: SEdgeNet, millimetre-wave radar, human activity recognition, point cloud, graph neural network, stochastic neighbour sampling, edge convolution, privacy-preserving sensing, embedded AI, ambient intelligence, depthwise separable convolution, assistive robotics