Davide Scaramuzza

We are thrilled to share our groundbreaking paper published today in @Nature: "Champion-Level Drone Racing using Deep Reinforcement Learning." We introduce "Swift," the first autonomous vision-based drone that beat human world champions in several fair head-to-head races! PDF (open access): nature.com/articles/s4158…. Video: youtu.be/fBiataDpGIo

The "Swift" AI drone combines deep #reinforcementlearning in simulation with data collected in the physical world. This marks the first time that an autonomous mobile robot has beaten human champions in a real physical sport designed for and by humans. The key difference to previous achievements, like IBM’s Deep Blue winning against Gary Kasparov at chess in 1996 or Google DeepMind’s AlphaGo crushing the top champion Lee Sedol at Go in 2016, is that the previous feats were limited to board games or video games.

This work represents a milestone for mobile robotics, machine intelligence, and beyond, which may inspire the deployment of hybrid learning-based solutions in other physical systems, such as autonomous vehicles, aircraft, and personal robots, across a broad range of applications.

Kudos to my coauthors @kaufmann_elia @l_bauersfeld @antoniloq, the pilots @captainvanover @bmsthomas, and our supporters @m_schaepman @UZH_en @UZH_Science @UZHspacehub @ERC_Research @nccrrobotics

We are excited to share our #CoRL2026 paper, "Learning Visuomotor Quadrotor Flight Through Dynamic Narrow Gaps via Differentiable Simulation", where we demonstrate autonomous quadrotor flight through narrow moving gaps using exclusively onboard vision, with policies trained from scratch in just minutes via #DifferentiableSimulation! Code released!
Paper: arxiv.org/pdf/2610.37557
Video: youtu.be/z-txORkhnrg
Code: github.com/uzh-rpg/agile_…

Flying through a narrow, moving gap using only egocentric vision is extremely challenging, as it requires centimeter-level precision, split-second timing, and perception-aware control. We use differentiable simulation to train vision-based control policies directly from scratch, without expert demonstrations, planners, or teacher policies. The key idea of uding differentiable simulation is that, instead of learning only from success or failure, the drone receives informative gradients that guide it toward better alignment, timing, and collision avoidance. Our approach achieves 99% success in simulation and 80% in the real world, with training completed in just minutes on a single GPU. Compared to PPO, it requires 20× fewer training transitions and 40× less training time! Real-world experiments use only onboard visual and inertial sensing. The motion capture visible in the video is used only for ground-truth measurements.

Kudos to Riccardo Bianco, Yunfan REN, and Jiaxu Xing!

@uzh_ifi @UZH_Science @UZH_en @UZHspacehub @ERC_Research @UZH_ai @AUTOASSESS_EU

I'm excited to announce that we have multiple fully funded positions in:
• Low-Energy, Low-Latency World Models with Event Cameras and Frames
• Vision-based Self-Improving Robot Learning in the Real World
• Spatial AI for Blind People

📅 Application deadline: October 27, 2026.
🔗 Details and application instructions: rpg.ifi.uzh.ch/positions.html

We are looking for highly motivated researchers who enjoy tackling fundamental questions and turning their ideas into working systems with real-world impact.

Through my affiliation as an associated faculty member of the @ETH_AI_Center, all members of my lab are automatically associated with the ETH AI Center.

Please share with students and researchers interested in joining us!

We are excited to share our paper "Temporal Cascading of Planning and Control for Quadrotor MPC," published in the IEEE Transactions on Robotics, where, instead of separating planning and control, we unify them in a single #ModelPredictiveController: a high-fidelity model controls the near future, while a low-fidelity model plans the long horizon!
PDF: arxiv.org/pdf/2512.12427
Narrated video: youtu.be/CefIqpld9N8

Traditionally, planning and control are treated separately: a simplified model plans a trajectory over a long horizon, and a high-fidelity controller then tries to track it. This is computationally convenient, but suboptimal: the controller is constrained by a plan that does not fully capture the drone dynamics.

We introduce a different way of looking at this problem. Instead of separating planning and control, we put them into one MPC optimization, but use different models at different points in time: a high-fidelity quadrotor model over the near future for accurate control, followed by a computationally cheap point-mass model over the longer horizon for planning. In other words: control precisely nearby, plan efficiently far ahead.

The key contribution is that the two horizons are not simply stacked: they are coupled through transition and feasibility constraints, so that what the simplified model plans remains compatible with what the real quadrotor can execute. We also use parallel low-fidelity planners and progressive obstacle smoothing to escape local minima in cluttered environments without slowing down the real-time control loop.

The result is long-horizon planning with iteration times below 5 ms, and in simulation and real-world flights, up to 75% lower closed-loop cost than standard MPC and conventional hierarchical planner-controller approaches under comparable computational budgets.

To recap: rather than choosing between an accurate short-horizon model and an efficient long-horizon model, use both at the time scales where each matters most!

Kudos to Rudolf Reiter, Chao Qin, Leonard Bauersfeld!

@uzh_ifi @UZH_Science @UZH_en @UZHspacehub @ERC_Research @UZH_ai @AUTOASSESS_EU

Three years ago, we celebrated @DanielGehrig6 ’s PhD defense. Today, we are delighted to celebrate his appointment as Assistant Professor at the @NUSingapore National University of Singapore, starting in March 2027, with the prestigious Presidential Young Professorship!

During his PhD in our lab at the University of Zurich, Daniel made pioneering contributions to event-based vision, from #TimeLens for video frame interpolation to our Nature paper, “Low Latency Automotive Vision with Event Cameras.” His work pushed the boundaries of fast, efficient visual perception for mobile devices and automated driving. His achievements also earned him the UZH Annual Award.

As his former PhD supervisor, it is a particular joy to see Daniel take this next step and establish his own research group!

Congratulations, Daniel! We are very proud of you and look forward to seeing what you and your future students will achieve!

Daniel’s homepage: danielgehrig18.github.io

UZH.ai @ERC_Research @uzh_ifi @UZH_Science @UZH_Science