Unmanned aerial vehicles (UAVs), commonly known as drones, are now widely used worldwide to tackle various real-world tasks, including filming videos for various purposes, monitoring crops or other environments from above, assessing disaster zones, and conducting military operations. Despite their widespread use, most existing drones either need to be fully or partly operated by human agents.

In addition, many drones are unable to navigate cluttered, crowded or unknown environments without colliding with nearby objects. Those that can navigate these environments typically rely on expensive or bulky components, such as advanced sensors, graphics processing units (GPUs) or wireless communication systems.

Researchers at Shanghai Jiao Tong University have recently introduced a new insect-inspired approach that could enable teams of multiple drones to autonomously navigate complex environments while moving at high speed. Their proposed approach, introduced in a paper published in Nature Machine Intelligence, relies on both a deep learning algorithm and core physics principles.

“Our research was inspired by the incredible flight capabilities of tiny insects like flies,” Prof. Danping Zou and Prof. Weiyao Lin, co-senior authors of the paper, told Tech Xplore. “It always amazed us how such small creatures, with only a tiny brain and limited sensing, can perform agile, intelligent maneuvers—avoiding obstacles, hovering mid-air, or chasing prey.

“Replicating that level of flight control has long been a dream and a major challenge in robotics. It requires tightly integrated perception, planning, and control—all running on very limited onboard computation, just like in the insect brain.”

Most common computational approaches for controlling the flight of multiple drones break down the task of autonomous navigation into separate modules, such as state estimation, mapping, path planning, trajectory generation and control modules. While tackling these sub-tasks separately can be effective, it sometimes prompts the accumulation of errors across different modules and introduces latency in the responses of drones. In other words, it can cause drones to react more slowly when they approach obstacles, which can increase the risk of collisions in dynamic and cluttered environments.

“The primary objective of our research was to explore whether a lightweight artificial neural network (ANN) could replace this classic pipeline with a compact, end-to-end policy,” said Prof. Zou and Prof. Lin.

“This network takes sensor data as input and directly outputs control actions—a paradigm that mirrors how flies use a small number of neurons to produce complex, intelligent behavior. We sought not just to match biological elegance, but to demonstrate that minimalism in sensing and computation can still yield high-performance autonomous flight.”

The new system developed by the researchers primarily relies on a newly developed lightweight artificial neural network that can generate control commands for a quadrotor aerial vehicle based on a 12×16 ultra-low-resolution depth map. While the definition of the maps fed to the algorithm is low, it was found to be sufficient for the network to make sense of its surrounding environment and effectively plan the actions of aerial vehicles.

“We trained this network in a custom-built simulator composed of simple geometric shapes—cubes, ellipsoids, cylinders, and planes—allowing us to generate diverse yet structured environments,” explained Prof. Zou and Prof. Lin. “Our training process is highly efficient, thanks to a differentiable physics-based pipeline. It supports both single-agent and multi-agent training modes: in the multi-agent setting, other drones are treated as dynamic obstacles during learning.”

A key advantage of the multi-aerial vehicle navigation approach developed by the researchers is that it relies on a highly compact and lightweight deep neural network that has only three convolutional layers. The researchers tested it on an embedded computing board that costs just $21 and found that it ran both smoothly and energy-efficiently.

“The training converges in just 2 hours on an RTX 4090 GPU, which is remarkably fast for policy learning,” said Prof. Zou and Prof. Lin. “Our system also naturally supports multi-robot navigation without any centralized planning or explicit communication, enabling scalable deployment in swarm scenarios.”

When they reviewed previous literature in the field, the researchers found that many deep learning algorithms for drone navigation did not generalize well across real-world scenarios. This is often because they do not account for unexpected obstacles or changes in the environment, and need to be trained with large amounts of flight data labeled by human experts.

“Our most important finding is that embedding the physics model of the quadrotor directly into the training process can significantly improve both training efficiency and real-world performance—in terms of robustness and agility,” said Prof. Zou and Prof. Lin.

“This technique, known as differentiable physics learning, wasn’t invented by us, but we are the first to extend and apply it successfully to real-world quadrotor control. Through this research, we also arrived at three unexpected yet powerful insights—lessons that could reshape how we think about intelligence, models, and perception in robotics.”

The promising findings attained by Prof. Zou, Prof. Lin and their colleagues demonstrate the potential of small artificial neural network-based models for tackling complex navigation tasks. The researchers showed that these models could be more effective than they are often perceived to be and can also help to understand how larger models work.

“Just as neuroscience made its early progress through the fruit fly, whose simple neural circuits helped unlock foundational insights, small models give us a clearer view of how perception, decision-making, and control are coupled,” said Prof. Zou and Prof. Lin. “In our case, a model with fewer than 2 MB of parameters enabled multi-agent coordination without any communication—showing how simplicity can lead to emergent intelligence.”

Notably, the lightweight model developed by the researchers performed well despite being trained in a simulated environment. This is in stark contrast to many previously developed models that require substantial amounts of expert-labeled data.

“We learned that intelligence doesn’t have to depend on massive datasets,” said the researchers. “We trained our policy entirely in simulation—without internet-scale data, pre-collected logs, or handcrafted demonstrations—using only a few basic tasks and geometric environments powered by a differentiable physics engine. This challenges the common assumption that ‘more data is always better’ and suggests that structural alignment and embedded physical priors may matter more than sheer data volume.”

Overall, the results of this recent study suggest that neural networks guided by basic physics principles could achieve better results than networks trained on millions of images, maps, or other labeled data. In addition, the researchers found that even a low definition depth image can be used to precisely guide the behavior of robots.

“Like the fruit fly, whose vision is limited to low-resolution compound eyes yet manages incredible aerial feats, we used 12×16-pixel depth images to control drones flying at speeds up to 20 m/s,” said Prof. Zou and Prof. Lin. “This supports a bold hypothesis: navigation performance may depend more on the agent’s internal understanding of the physical world than on sensor fidelity alone.”

In the future, the approach developed by Prof. Zou, Prof. Lin and his colleagues could be deployed on more types of aerial vehicles and tested in specific real-world scenarios. Eventually, it could help to broaden the tasks that can be tackled by ultra-lightweight drones, for instance, allowing them to automatically take selfies or compete in racing competitions. The approach could also prove useful for broadcasting sports or other events, for searching collapsed buildings during search and rescue operations, and for inspecting of cluttered warehouses.

“We are currently exploring the use of optical flow instead of depth maps for fully autonomous flight,” added Prof. Zou and Prof. Lin. “Optical flow provides fundamental motion cues and has long been studied in neuroscience as a key component of insect vision.

“By using it, we hope to get even closer to mimicking the natural strategies that insects use for navigation. Another important direction we’re pursuing is the interpretability of end-to-end learning systems.”

Although the team’s lightweight neural network was found to perform remarkably well in real-world experiments, how these promising results work is not yet fully understood. As part of their next studies, Prof. Zou and Prof. Lin hope to shed more light on the network’s internal representations, which could also offer insights into how insects process their surroundings and plan their actions.

© 2025 Science X Network