The standard approach to making autonomous robots safe – adding rules and constraints – may be the wrong approach entirely. A new drone racing paper suggests letting safety emerge from competition instead.
When engineers want a robot to behave safely around other moving things, they write safety rules into its software. Avoid this zone. Slow down near obstacles. Keep a minimum separation. The constraints stack up, and so does the fragility: edge cases break rules, and the real world is full of edge cases.
A paper posted to arXiv on 21 May by researchers at the University of Zurich and Google DeepMind suggests a different path. Instead of constraining a single drone with rules, train a group of drones to race against each other – and let safety emerge from the competition itself.
What the researchers did
The team used high-speed quadrotor racing as a testing ground. Quadrotors are the four-propeller drones common in photography and delivery. At racing speeds above 22 metres per second, roughly 80 kilometres per hour, the aerodynamics become treacherous. When one drone flies below another, turbulence from its rotors, called downwash, destabilises the drone above it. Human pilots manage this instinctively. AI systems trained alone against static tracks do not.
The researchers trained their drones using multi-agent reinforcement learning. In reinforcement learning, a system learns by trial and error, rewarded for good outcomes and not for bad ones. Training multiple drones simultaneously against each other forced them to learn to overtake, block, avoid collisions, and anticipate what opponents would do.
No explicit collision-avoidance rules were programmed in. The drones learned to avoid each other because crashing was bad for winning.
What they found
The resulting agents beat a champion-level human pilot in multi-drone races. More significantly, they cut collision rates by 50 per cent compared to the best existing single-agent systems – systems already capable of avoiding static obstacles, but brittle when other moving drones appeared.
Crucially, the AI had never trained against real human pilots. When tested against them anyway, it adapted successfully, giving way when appropriate and competing when it could. No additional training was needed.
Why this matters
The conventional assumption in autonomous systems design is that safety comes from constraints on individual agents. This paper challenges that at a practical level. A robot trained in isolation develops robustness to fixed environments. A robot trained against unpredictable opponents develops robustness to unpredictable agents – which is the actual problem in every real-world deployment.
Autonomous vehicles share roads with human drivers. Delivery robots navigate streets full of pedestrians. Warehouse robots work beside human pickers. In every case, the engineering problem is not “navigate a fixed obstacle course” but “navigate around agents with their own intentions.” Multi-agent training addresses that directly.
There is also a specific implication for commercial drone airspace. Urban drone delivery networks will eventually involve multiple autonomous aircraft operating in proximity, and managing aerodynamic interactions between them is unsolved. A control framework that learns to handle downwash without explicit aerodynamic modelling is a meaningful step forward.
What comes next
This is a preprint, not yet peer-reviewed. The human-pilot tests involved one champion competitor in a controlled racing environment, not open airspace. Extending the approach to public airspace involves regulatory and coordination challenges the paper does not address.
But the core result stands. Training drones to race each other produced safer drones than training them to avoid obstacles. That is a direct challenge to how autonomous systems are currently built, and the implications extend well beyond racing.
Sources
Geles, I., Bauersfeld, L., Wulfmeier, M., & Scaramuzza, D. (2026). Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning. arXiv preprint arXiv:2605.22748. Posted 21 May 2026. https://arxiv.org/abs/2605.22748
University of Zurich, Robotics and Perception Group — Ismail Geles, Leonard Bauersfeld, Davide Scaramuzza https://rpg.ifi.uzh.ch/marl (project page, linked in the preprint)
Project website with video: https://rpg.ifi.uzh.ch/marl
Quotes in this article are drawn from the preprint “Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning” by Ismail Geles, Leonard Bauersfeld, Markus Wulfmeier, and Davide Scaramuzza, posted to arXiv (arXiv:2605.22748) on 21 May 2026.
Ray Jackson holds a BSc in Electrical Engineering from the University of Manitoba and a PhD in Physics from Carleton University. His reporting interests include Current and Future Technologies, Engineering and Artificial Intelligence.