Skip to main content
Back

AI Evolves Bizarre Asymmetric Drones That Beat Traditional Hexacopters by 170%

Researchers evolved asymmetric hexacopter drones that beat conventional designs by up to 170%, using evolutionary algorithms to shape bodies and PPO to train brains.

Pro

Paper, Full Access
  • Evolved asymmetric hexacopters outperformed the standard design by 29% to 170% across four navigation tasks.
  • Bodies were evolved with a (μ+λ) mutation-only strategy while controllers were trained fresh with PPO per morphology.
  • Best gains came on shuttle run (+170%) and figure8 (+143%), tasks requiring aggressive direction changes.
  • Evolution actively selected against central symmetry, producing designs that look nothing like human-engineered drones.
  • New domain-agnostic metrics (burn-in, learning speed, volatility) reveal evolved bodies are progressively easier to train.
  • Work stays in simulation; no code release and real-world transfer left as future work.

What if the reason your drone looks the way it does is not because that shape is optimal, but because a human engineer thought symmetry looked right? A team from Vrije Universiteit Amsterdam and TU Delft removed human bias from drone design, letting an evolutionary algorithm invent hexacopter bodies while reinforcement learning taught each new body how to fly. The resulting drones look bizarre and asymmetric, and they beat the standard hexacopter on every task tested.

The paper tackles a niche but important question in embodied AI: when you can co-optimize a robot's body and its controller, what does the machine come up with? Decoupling design from human intuition should open up a broader array of viable body plans.

Splitting the drone into a body problem and a brain problem

The setup follows what the authors call the Triangle of Life framework, which treats a robot as a body (morphology) plus a brain (controller). Morphologies are evolved, while controllers are learned from scratch through reinforcement learning for each new body.

Diagram comparing evolved drone body and learned neural network controller

Each drone is encoded as 36 parameters covering the six arms: arm length, two rotation angles for the arm itself, two more for the motor at the end, and a binary flag for spin direction. Initial populations are drawn randomly, but naive sampling produces many drones that cannot generate enough thrust in a single direction to overcome gravity. A static hovering feasibility test filters those out before any compute is wasted training them.

Evolution runs as a mutation-only (μ+λ) strategy with a population of 12 and 40 generations. Crossover was dropped because small genotype changes produced wildly different phenotypes, breaking the locality property evolutionary algorithms rely on. A repair operator pushes overlapping rotors apart to keep every design physically buildable.