Drones and Aerial Robotics

Aerial robotics is the design and control of robots that fly, from delivery drones and racing quadrotors to swarmsswarm roboticsSwarm robotics is coordinating many relatively simple robots through local interaction, without a central planner, for collective behavior no individual achieves, such as coverage, mapping or search at fleet scale. Classical models of pairwise attractions and repulsions (potential fields, after Reynolds flocking) explain collective motion but guarantee neither safety nor speed in clutter. Aerial swarms now optimize trajectories onboard: each drone plans in milliseconds from its own sensors, treating neighbors as constraints. Ten drones crossed a bamboo forest this way with no external localization or global map.. Flight was an early learning success because its dynamics simulate well: a reinforcement-learning system beat human world champions in its own lab in 2023^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023, and a one-camera drone won an organiser-run race in 2025^{Source 9}9 Bahnam et al. 2026, MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI. Stavrow A. Bahnam, Robin Ferede, Till M. Blaha et al., 2026.

Photo: Commanderbryce / Wikimedia Commons. Licence: CC BY-SA 4.0.

Switch the drone below between an ordinary camera, which takes about 70 ms to report what it sees, and a faster event camera, and try each swerve setting: the top safe speed moves with every choice^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019.

Why flight was the first solved testbed

A quadrotor is a benign case for simulation. Its dynamics are differentially flat in position and heading, even with linear rotor drag^{Source 10}10 Faessler et al. 2018, Differential Flatness of Quadrotor Dynamics Subject to Rotor Drag for Accurate Tracking of High-Speed Trajectories. Matthias Faessler, Antonio Franchi, Davide Scaramuzza, IEEE Robotics and Automation Letters 3(2):620-626, 2018, and standard models cover its dynamics, estimation and control^{Source 11}11 Mahony et al. 2012, Multirotor Aerial Vehicles: Modeling, Estimation, and Control of Quadrotor. Robert Mahony, Vijay Kumar, Peter Corke, IEEE Robotics & Automation Magazine 19(3):20-32, 2012. Flight also has no contact: an aerial robot either flies in free space or has already failed. Racing quadrotors still exceed 100 km/h and accelerate at several times gravity^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023. Make the drone in the figure below climb, fly forward and turn, and watch its four propeller speeds change.

How does a drone steer with no wings or rudder?

Four motors, a battery and a flight computer. Seen from above, two propellers spin clockwise and two counter-clockwise.

BackNext

A drone has no wings or rudder, so it moves and turns only by changing how fast each propeller spins.

How this was made

The steps in words

1. Four motors, a battery and a flight computer. Seen from above, two propellers spin clockwise and two counter-clockwise.
2. Spin all four at the same speed and it hovers: their upward push matches its weight. Faster together, it climbs.
3. Now fly forward. No propeller points forward, so something else has to push it that way.
4. Each propeller twists the drone against its spin. Speed up the clockwise pair and their twist wins: the drone turns counter-clockwise.
5. A gust tips it. The flight computer feels the tilt and resets the four speeds, often 400 times a second. This is attitude control.

Parts

• Frame
• Motors
• Propellers
• Battery
• Flight computer

Classical stacks plan a trajectory and then track it. Mellinger and Kumar generated minimum-snap trajectories for quadrotors^{Source 12}12 Mellinger et al. 2011, Minimum snap trajectory generation and control for quadrotors. Daniel Mellinger, Vijay Kumar, IEEE ICRA 2011, 2520-2525, and Lee and colleagues built a geometric tracking controller on SE(3)^{Source 13}13 Lee et al. 2010, Geometric tracking control of a quadrotor UAV on SE(3). Taeyoung Lee, Melvin Leok, N. Harris McClamroch, 49th IEEE Conference on Decision and Control (CDC), 5420-5425, 2010. In 2021 time-optimal trajectory planning outperformed professional human pilots in racing^{Source 14}14 Foehn et al. 2021, Time-optimal planning for quadrotor waypoint flight. Philipp Foehn, Angel Romero, Davide Scaramuzza, Science Robotics 6(56):eabh1221, 2021.

Swift, from the University of Zurich, raced three human champions, including the world champions of two leagues, and won several races against each^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023. It trained its control policy in simulation with model-free reinforcement learning^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023. The team then fitted residual models of perception noise and unmodeled dynamics from real flight data and retrained^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023.

In September 2023, Song and colleagues showed that an RL controller outperformed optimal control on the same racing task^{Source 3}3 Song et al. 2023, Reaching the Limit in Autonomous Racing: Optimal Control versus Reinforcement Learning. Yunlong Song, Angel Romero, Matthias Mueller et al., Science Robotics 2023, adg1462. Optimal control splits the problem into planning and control with a trajectory as the interface, which limits the controller under unmodeled effects. RL optimizes a task-level objective directly and handles model uncertainty through domain randomization; it flew above 12 g and 108 km/h^{Source 3}3 Song et al. 2023, Reaching the Limit in Autonomous Racing: Optimal Control versus Reinforcement Learning. Yunlong Song, Angel Romero, Matthias Mueller et al., Science Robotics 2023, adg1462. A later hybrid embedded a differentiable MPC in an actor-critic framework and matched model-free RL at up to 21 m/s^{Source 15}15 Romero et al. 2026, Actor-Critic Model Predictive Control: Differentiable Optimization Meets Reinforcement Learning for Agile Flight. Angel Romero, Elie Aljalbout, Yunlong Song et al., IEEE Transactions on Robotics 42:673-692, 2026.

In April 2025 a TU Delft drone with one rolling-shutter camera and an IMU won the A2RL race in Abu Dhabi, beating all AI teams and three human world champions in a knockout tournament^{Source 9}9 Bahnam et al. 2026, MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI. Stavrow A. Bahnam, Robin Ferede, Till M. Blaha et al., 2026. TU Delft notes that Swift raced in a lab where the researchers controlled conditions, hardware and track, while the A2RL organisers designed both hardware and track^{Source 16}16 Delft 2025, Autonomous drone from TU Delft defeats human champions in historic racing first. TU Delft, TU Delft news, 2025-04-15.

In January 2026 world FPV champion Minchan Kim beat TII Racing's drone from a 4 to 4 tie when it struck a gate and could not recover^{Source 17}17 A2RL 2026, A2RL Drone Championship Sets the Pace for AI in Autonomous Flight. A2RL, A2RL news, 2026-01-23. Champion policies stay narrow: they often crash immediately in unseen configurations, and a 2026 method improved generalization 7.4x^{Source 18}18 Green et al. 2026, Bridging Performance and Generalization in Reinforcement Learning for Agile Flight. Jonathan Green, Jiaxu Xing, Nico Messikommer et al., 2026.

The autonomy stack on a flying robot

One camera and one low-cost IMU are the minimum sensor suite for metric six-DoF state estimation^{Source 19}19 Qin et al. 2018, VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. Tong Qin, Peiliang Li, Shaojie Shen, IEEE Transactions on Robotics 34(4):1004-1020, 2018. Visual-inertial odometryvisual-inertial odometry (VIO)Visual-inertial odometry (VIO) is state estimation for a flying robot fusing camera images with inertial measurement unit data. The camera fixes long-term drift, the IMU gives high-rate accelerations between frames, and a filter or optimizer over both yields the metric pose and velocity flight control needs. Needing no external infrastructure or GPS, it is the standard onboard localization for small drones. Swift's perception fuses a visual-inertial estimator and a convolutional gate detector in a Kalman filter for its control policy. fuses image features with IMU measurements into the pose and velocity the controller needs, without GPS or motion capture^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023. The fast inner loop often runs on a microcontroller stack such as PX4, which has substantially lower latency than a typical robotics Linux system^{Source 20}20 Meier et al. 2015, PX4: A node-based multithreaded open source robotics framework for deeply embedded platforms. Lorenz Meier, Dominik Honegger, Marc Pollefeys, IEEE ICRA 2015, 6235-6240.

Photo: Sinky / Wikimedia Commons. Licence: CC BY-SA 4.0.

Perception latency bounds the whole system

Falanga, Kim and Scaramuzza asked how fast a drone can fly given its perception latency and the acceleration its airframe can produce^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019. A drone at speed $v$ sees an obstacle at sensing range $s$, and its pipeline delays that information by latency $\tau$. Avoiding the obstacle at maximum lateral acceleration $u$ takes time proportional to the square root of obstacle size over $u$. If $s/v - \tau$ is shorter than the maneuver, no controller saves the drone^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019.

A 70 ms stereo pipeline, the upper bound the study took from a commercial camera's datasheet, supports about 19 m/s at 25 m/s² of lateral acceleration and 25 m/s at 50 m/s²^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019. An event camera reports per-pixel brightness changes without waiting for a frame. At 8 m of range its ceiling is 7 percent higher on the least agile airframe and 22 percent higher at 50 m/s², 31.03 against 25.40 m/s. The paper's own summary of 7 to 12 percent is for its most agile airframe, at 200 m/s², against the fastest stereo camera^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019.

In 2020 an event-camera quadrotor dodged fast-moving objects using only onboard sensing and computation^{Source 21}21 Falanga et al. 2020, Dynamic obstacle avoidance for quadrotors with event cameras. Davide Falanga, Kevin Kleber, Davide Scaramuzza, Science Robotics 5(40):eaaz9712, 2020. Vision-language-action flight is far slower: RaceVLA averaged 1.04 m/s with a top speed of 2.02 m/s^{Source 22}22 Serpiva et al. 2025, RaceVLA: VLA-based Racing Drone Navigation with Human-like Behaviour. Valerii Serpiva, Artem Lykov, Artyom Myshlyaev et al., 2025. On the CognitiveDrone benchmark RaceVLA succeeded 31.3 percent of the time, and CognitiveDrone-R1, which adds a reasoning module, 77.2 percent^{Source 23}23 Lykov et al. 2025, CognitiveDrone: A VLA Model and Evaluation Benchmark for Real-Time Cognitive Task Solving and Reasoning in UAVs. Artem Lykov, Valerii Serpiva, Muhammad Haris Khan et al., 2025.

Jaewoo Lee, Guanya Shi and colleagues at Seoul National University, Carnegie Mellon and elsewhere (October 2026) let a coding agent that programs an aerial manipulator run diagnostic programs in simulation, such as a camera move or a test lift, before it declares a task done^{Source 37}37 Lee et al. 2026, Fly-by-Code: Embodied Coding Agents for Aerial Manipulation with Active Visual and Physical Feedback. Jaewoo Lee, Jeongyeon Seo, Sihyun Cho et al., 2026. Over six simulated tasks and 60 trials per condition, the authors report 75 percent success against 31.7 percent with execution logs and final images^{Source 37}37 Lee et al. 2026, Fly-by-Code: Embodied Coding Agents for Aerial Manipulation with Active Visual and Physical Feedback. Jaewoo Lee, Jeongyeon Seo, Sihyun Cho et al., 2026. Unmodified, the programs completed 4 of 5 real two-object pick-and-place trials and 3 of 5 cabinet trials^{Source 37}37 Lee et al. 2026, Fly-by-Code: Embodied Coding Agents for Aerial Manipulation with Active Visual and Physical Feedback. Jaewoo Lee, Jeongyeon Seo, Sihyun Cho et al., 2026.

End-to-end learning skips the interfaces

Deep Drone Acrobatics trained a policy entirely in simulation by imitating an optimal controller with privileged information, and flew maneuvers at up to 3 g^{Source 24}24 Kaufmann et al. 2020, Deep Drone Acrobatics. Elia Kaufmann, Antonio Loquercio, René Ranftl et al., RSS 2020. In 2021 Loquercio and colleagues used the same recipe to map noisy depth images directly to collision-free trajectories^{Source 1}1 Loquercio et al. 2021, Learning High-Speed Flight in the Wild. Antonio Loquercio, Elia Kaufmann, René Ranftl et al., Science Robotics 2021, Vol. 6, Issue 59, abg5810. With realistic simulated sensor noise, the policy transferred to forests, snow, derailed trains and collapsed buildings with no additional training. The authors argue that separating sensing, mapping and planning "can be problematic for high-speed navigation in cluttered environments" because latency compounds and errors propagate^{Source 1}1 Loquercio et al. 2021, Learning High-Speed Flight in the Wild. Antonio Loquercio, Elia Kaufmann, René Ranftl et al., Science Robotics 2021, Vol. 6, Issue 59, abg5810.

Swift kept a modular front end: a visual-inertial estimator and a gate-detection network feed a Kalman filter, and a small two-layer RL network maps that state to collective thrust and body rates^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023. Later systems drop explicit state estimation. Geles and colleagues fly from pixels with the thrust and body-rate commands human pilots use^{Source 25}25 Geles et al. 2024, Demonstrating Agile Flight from Pixels without State Estimation. Ismail Geles, Leonard Bauersfeld, Angel Romero et al., Robotics: Science and Systems XX, 2024, and Dream to Fly learned racing from pixels with a world model where PPO and SAC struggled^{Source 26}26 Romero et al. 2025, Dream to Fly: Model-Based Reinforcement Learning for Vision-Based Drone Flight. Angel Romero, Ashwin Shenai, Ismail Geles et al., ICRA 2026, 2025.

MonoRace keeps a state estimator, but its network skips inner-loop controllers and sends motor commands at 500 Hz^{Source 9}9 Bahnam et al. 2026, MonoRace: Winning Champion-Level Drone Racing with Robust Monocular AI. Stavrow A. Bahnam, Robin Ferede, Till M. Blaha et al., 2026. Training through differentiable physics produced a depth-to-control policy that flies forests at up to 20 m/s on a $21 computer^{Source 27}27 Zhang et al. 2025, Learning vision-based agile flight via differentiable physics. Yuang Zhang, Yu Hu, Yunlong Song et al., Nature Machine Intelligence 7(6):954-966, 2025.

Swarm robotics

Swarm robotics uses many cheap platforms to cover a search area, map a forest or fly a light show. Its founding models came from biology, Reynolds' 1987 flocking model being the canonical one^{Source 8}8 Reynolds 1987, Flocks, Herds and Schools: A Distributed Behavioral Model. Craig W. Reynolds, Computer Graphics 21(4):25-34 (SIGGRAPH '87), 1987, and treat the group as a dynamical system of local attraction and repulsion. Vásárhelyi and colleagues optimized those terms and flew 30 drones in coherent outdoor flocking^{Source 7}7 Vásárhelyi et al. 2018, Optimized flocking of autonomous drones in confined environments. Gábor Vásárhelyi, Csaba Virágh, Gergő Somorjai et al., Science Robotics 3(20), eaat3536, 2018. Reactive models have no lookahead.

Soria, Schiano and Floreano folded the interaction terms into a nonlinear model-predictive controller with obstacle and neighbour terms^{Source 6}6 Soria et al. 2021, Predictive control of aerial swarms in cluttered environments. Enrica Soria, Fabrizio Schiano, Dario Floreano, Nature Machine Intelligence 3, 545-554, 2021. In simulation it flew faster and with fewer collisions than the reactive equivalent, and five quadrotors flew it through a real indoor obstacle field^{Source 6}6 Soria et al. 2021, Predictive control of aerial swarms in cluttered environments. Enrica Soria, Fabrizio Schiano, Dario Floreano, Nature Machine Intelligence 3, 545-554, 2021. EGO-Swarm's agents generate safe, dynamically feasible trajectories in several milliseconds over an unreliable trajectory-sharing network^{Source 28}28 Zhou et al. 2020, EGO-Swarm: A Fully Autonomous and Decentralized Quadrotor Swarm System in Cluttered Environments. Xin Zhou, Jiangchao Zhu, Hongyu Zhou et al., ICRA 2021, 2020.

Zhou and colleagues ran a spatial-temporal trajectory optimizer on palm-sized drones, solving in milliseconds from onboard sensing with no external localization^{Source 5}5 Zhou et al. 2022, Swarm of micro flying robots in the wild. Xin Zhou, Xiangyong Wen, Zhepei Wang et al., Science Robotics 7(66), eabm5954, 2022. Ten of them flew through a dense bamboo forest^{Source 5}5 Zhou et al. 2022, Swarm of micro flying robots in the wild. Xin Zhou, Xiangyong Wen, Zhepei Wang et al., Science Robotics 7(66), eabm5954, 2022. The differentiable-physics policy coordinates several drones without communication or centralized planning^{Source 27}27 Zhang et al. 2025, Learning vision-based agile flight via differentiable physics. Yuang Zhang, Yu Hu, Yunlong Song et al., Nature Machine Intelligence 7(6):954-966, 2025. Multi-agent RL with league self-play beat a champion-level human pilot in multi-player races above 22 m/s and cut collisions by 50 percent against single-agent baselines^{Source 29}29 Geles et al. 2026, Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning. Ismail Geles, Leonard Bauersfeld, Markus Wulfmeier et al., 2026.

Everything cited above coordinates implicitly; none of it decides who goes where. In Ukraine, electronic warfare is pushing military drones toward terrain-based navigation, autonomous target lock-on and, eventually, mesh-networked swarms^{Source 30}30 Pultarova 2026, How Autonomous Drone Warfare Is Emerging in Ukraine. Tereza Pultarova, IEEE Spectrum, 2026-03-24.

Delivery and regulation

Zipline reports more than 3 million deliveries and more than 140 million commercial autonomous miles since its first commercial delivery in 2016^{Source 31}31 Zipline n.d., Safety Facts. Zipline, Zipline, n.d.; accessed 2026-10-14. In January 2026 it reported US delivery growth of about 15 percent week over week for seven months^{Source 32}32 Zipline 2026, Zipline Surpasses 2 Million Deliveries, Raises More than $600M to Power Next Phase of Growth, and Expands Operations to Houston and Phoenix. Team Zipline, Zipline newsroom, 2026-01-21. Wing reports well over one million commercial deliveries and plans more than 270 Walmart locations reaching 40 million Americans by 2027^{Source 33}33 Wing 2026, Wing and Walmart name seven new markets for drone delivery service. Wing Team, Wing news, 2026-06-08. Walmart reported 1 million drone deliveries in May 2026, from 66 stores in four states^{Source 34}34 Walmart 2026, Walmart Celebrates 1 Million Drone Deliveries, Marking a Major Milestone in Customer Convenience. Walmart, Walmart corporate news, 2026-05-29.

The FAA and TSA proposed Part 108 in August 2025 to enable routine flight beyond visual line of sight^{Source 35}35 FAA et al. 2025, Normalizing Unmanned Aircraft Systems Beyond Visual Line of Sight Operations. Federal Aviation Administration, Transportation Security Administration, Federal Register 2025-14992 (NPRM), 2025-08-07. The FAA reopened comments on electronic conspicuity and right-of-way until February 11, 2026^{Source 36}36 FAA 2026, Normalizing Unmanned Aircraft Systems Beyond Visual Line of Sight Operations; Reopening of Comment Period. Federal Aviation Administration, Federal Register 2026-01644, 2026-01-28.

What the rest of robotics takes from it

Latency budgets transfer directly. A policy that infers at 20 Hz controls a contact-rich task at about 20 Hz whatever the arm's bandwidth, the reasoning the perception-latency study formalised for flight^{Source 4}4 Falanga et al. 2019, How Fast Is Too Fast? The Role of Perception Latency in High-Speed Sense and Avoid. Davide Falanga, Suseong Kim, Davide Scaramuzza, IEEE Robotics and Automation Letters 4(2), 1884-1891, 2019. Real-time execution covers that budget for manipulators.

The sim-to-real recipe also carries over: residual models fitted from real data^{Source 2}2 Kaufmann et al. 2023, Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio et al., Nature 620, 982-987, 2023 plus privileged imitation with realistic sensor noise^{Source 1}1 Loquercio et al. 2021, Learning High-Speed Flight in the Wild. Antonio Loquercio, Elia Kaufmann, René Ranftl et al., Science Robotics 2021, Vol. 6, Issue 59, abg5810. Sim-to-real transfer and parallel simulation cover it for other robots.

Champion claims need context. Swift won in its own lab and MonoRace on organiser hardware^{Source 16}16 Delft 2025, Autonomous drone from TU Delft defeats human champions in historic racing first. TU Delft, TU Delft news, 2025-04-15, and the 2026 final turned on one unrecovered gate strike^{Source 17}17 A2RL 2026, A2RL Drone Championship Sets the Pace for AI in Autonomous Flight. A2RL, A2RL news, 2026-01-23. The distance between peak and reliable performance is the reliability gap, and surgical robotics shows a bar cleared under regulatory oversight.