The need to manage large fleets of autonomous vehicles and drones continues to grow. Unfortunately, the systems in place today require a lot of effort and computational power. Additionally, these safety programming systems cannot adjust on the fly.
Consequently, as commercial use of drone fleets becomes common, it’s vital to create safety programming protocols that streamline the processes and provide guarantees. Thankfully, a team of innovative researchers from MIT may have a solution. Here’s what you need to know.
Drone Shows
From the Olympics to new product releases, drone shows have become more common than ever. These events can utilize hundreds of drones that work in unison to create images, animations, and much more. These large-scale light displays look breathtaking. Even more impressive than these actions is all the work that goes into making these shows possible.
Current Safety Programming Methods
The current way that engineers control large drone swarms or other multi-autonomous robotic systems is through the use of Multi-agent systems (MAS). These protocols combine trajectories, with waypoints and time constraints. Together, these factors allow each drone to know where it is along its pre-planned flight path.
This approach works when conditions are good, and the drones are where they need to be on time. However, some serious drawbacks can occur when devices go off course. To prevent collisions, engineers employ a technique called pair-wise path-planning.
Safety Programming Drawbacks
This algorithm lets each drone know where it’s supposed to be relative to other drones if everything is going perfectly. However, when there are miscommunications, severe problems can arise. Since each drone in the MAS setup is flying blind after launch, it will presume the preplanned course even if outside factors make the original plan incorrect. This situation can lead to some very strange actions, like drones landing unexpectedly, falling from the sky onto onlookers, flying off into nowhere, or becoming unresponsive.
Recent Drone Show Accidents
The sudden rise in drone shows and autonomous robotics has led to some hazardous situations. Drones pose many threats, even when unarmed. These devices can lose connection and fall from the sky from hundreds of feet, damaging the device and pedestrians. Additionally, collisions with other aircraft could result in catastrophic losses and death. Two recent examples highlight these concerns:
Florida
The Orlando residents who gathered outside to watch the drone show on December 21, 2024, had no idea the event would turn dangerous. During the event, which included hundreds of drones, a connection error resulted in several drones plummeting to the ground and striking viewers.
Sadly, one of the drones fell onto a 7-year-old child, causing severe injuries that required emergency surgery. When questioned about the incident, the company behind the event, Sky Elements Drone, expressed remorse and pledged to investigate the cause to prevent future incidents.
New York
Another incident that occurred last year involved drones in Staten Island, New York. In this incident, a drone operator had flown 2.5 miles away. Being out of the clear line of sight, he was unable to notice a UH-60 Black Hawk helicopter and collided with the craft.
Thankfully, the helicopter was able to limp back to base. However, under further inspection, the crew realized that they had gotten extremely lucky as pieces from the drone were discovered in the main rotor shaft. After a lengthy investigation, the National Transportation Safety Board issued an accident report that revealed the UAV operator’s fault.
Multi-Agent Safety Programming Study
Recognizing the need for a more agile and responsive safety system when dealing with fleets of autonomous units, MIT engineers introduced a new training method in a recent paper published in IEEE Transactions on Robotics1. Their research details a new training method that enhances safety protections and reduces the workload surrounding insulting these protocols.
Graph Control Barrier Function GCBF+
The Graph Control Barrier Function + protocol builds on earlier barrier function algorithms and adds proprietary systems to streamline maintaining safety across massive fleets of vehicles. Impressively, GCBF+ allows drones to navigate complex environments much like people do.
Sensing Radius
The system begins by creating a smaller sensing radius that focuses mainly on collision risks rather than the entire flight plan. This reduces the drone’s management needs and allows it to actively track and avoid obstacles in the surrounding areas.
Situational Awareness
The updated GCBF+ allows drones to know exactly where they are in real-time and their positioning concerning other UAVs in the vicinity. Tracking all the drones in real-time enables multiple coordinated, collaborative, and computer-programmed agents to work together to accomplish tasks.
Graph Neural Networks (GNNs)
The team created a custom computer model that leveraged advanced Graph Neural Networks, enabling them to take advantage of some key benefits like the ability to parameterize a GCBF and distribute control policies. Specifically, the system simulates agents and controllers.
Source – MIT
The engineers utilize the exact specifications of real-world drones. They included their mechanical capabilities, limits, performance, battery life, and other crucial factors. The simulation then took this information and used it to create large-scale tests that engineers monitored.
Notably, the decision to use a graph neural network (GNN) makes sense as it enables tracking of the changing graphical topology of distance-based observation information flow. This input was then used in conjunction with other data to program the AI system to balance out performance and safety.
Add More Agents and Tasks
The engineers then began to scale up the simulation to include more units and more complex tasks. They noted that the real-time safety systems automatically adjusted after being copied and pasted to new drones. Each interaction was noted as more drones were added to the simulation.
Track Collision Over Thousands of Simulations
As the number and density of agents and obstacles increased, the system was fine-tuned to see no collisions as a reward. The GNN then autonomously began adjusting the controller input to ensure that safety violations were reduced.
The engineers took note of how the GNN enabled the drones to adjust their trajectories in real-time to avoid collisions with other bots. This test helped to ensure that their controller was reactive, meaning that it was constantly recreating a flight path for the drones based on real-time environmental conditions.
The MIT engineers also introduce safety boundaries as part of this approach. These are areas where devices are likely to run into safety violations. By instructing devices to avoid these areas, the team drastically reduced many of the collisions and errors plaguing the current systems.
Training Method for Multi-Agent Systems
Impressively, the controller was copied and pasted to other units once the simulation worked out any safety concerns. Since the protocol can handle arbitrary graph topologies, it’s easier to scale vertically. Additionally, it can account for changing numbers of participants, meaning that programming can be done on a few drones and then shared with the swarm.
MAS Safety Programming Test
The engineers set off to test their distributed framework for safe multi-agent control in large-scale environments with obstacles through the use of Crazyflies drones. These hand-sized quadrotor drones were upgraded using the GNN. In total 8 Crazyflies were used in the testing phase.
The agents used the 8 real-world drones to track and monitor activities. Then, the data was shared in a simulation with 1000 drones. The same scale-up method was applied for obstacles. The team began with 8 obstacles in a real-world test. The simulation increased the amount to 128 moving and non-moving obstacles for the final tests.
MAS Safety Programming Test Different Objectives
The drones were provided with different objectives to see how they could fare in crowded environments. One task was to switch positions while in flight. This task may seem simple but when you are dealing with 1000 drones, the risk of collision is high.
Landing
The next objective was to have drones land on moving boxes. The moving boxes were robots called Turtlebots. The Turtlebots were set up to circle at different speeds. The drones needed to navigate around each other and safely land on the turtlebot to complete their task.
MAS Safety Programming Test Results
The results of the safety programming study are eye-opening. For one, the devices outperformed their predecessor in both the 2D and 3D environments. The system’s use of LiDAR-based point-cloud observations to handle obstacles proved a great addition. It enabled the drones to make real-time adjustments and remain in their designated safety zones.
The Crazyflies flew around, completed their tasks mid-flight, and landed successfully without collision. The engineers concluded that their system could provide a 40% reduction in collisions. It also improved drone performance, allowing them to complete tasks that involved hundreds of drones without incident.
MAS Safety Programming Study Benefits
There are several benefits that this study brings to the market. For one, it allows engineers to automatically balance satisfying safety requirements and performance criteria. The system’s use of control barrier function theory for safety guarantees reduces hazards to those in the vicinity of these crafts dramatically. Also, the use of LiDAR systems ensures real-time feedback and flight path adjustments.
Destination Only
One of the biggest advantages of this technology is that engineers didn’t plan out any flight path versus traditional methods that require a lot of time and computational effort to create. The team only gave the drone instructions and a destination. The device didn’t compute a single collision-free trajectory. Instead, it computed thousands of trajectories based on its sensory input in real-time, ensuring that all changes are tracked and adjustments made.
Scalable
The engineers made a major upgrade to today’s system with their latest copy-and-paste controller. This protocol utilizes a graph structure that is ideal for scalable and generalizable distributed control of MAS.
This approach means that future drone pilots will only need to program a small number of units. The safety parameters can then be scaled up to an unlimited amount of drones, saving money, time, and effort, while not sacrificing performance.
MAS Safety Programming Applications
The list of applications for this technology continues to grow. The ability to easily create, distribute, and scale the controller protocol is a huge upgrade to drone anti-collision systems. It allows for large-scale applications to be quickly configured for new environments. This capability makes this tech suitable for use in warehouses, search and rescue operations, self-driving cars, and military tasks.
MAS Safety Programming Researchers
The researchers behind this task include MI Associate Professor of Aeronautics and Astronautics, Chuchu Fan. Additionally, Songyuan Zhang, Oswin, and Kunal Garg assisted in the research. Notably, the study received financial support from the U.S. National Science Foundation, MIT Lincoln Laboratory under the Safety in Aerobatic Flight Regimes (SAFR) program, and the Defence Science and Technology Agency of Singapore.
A Leading Company in Autonomous Vehicles
Several companies could leverage the MAS safety programming study to improve their UAV or EV offerings. As more autonomous craft manufacturers launch, the safety requirements for fleets of autonomous vehicles will increase. Here is one company that’s positioned perfectly to leverage this data and improve its ROIs.
When it comes to operating swarms of autonomous robots, Amazon (AMZN -0.66%) is the leader of the pack. The company integrated autonomous robots into its factory many years ago. Since then it has expanded its drone fleet to include UAVs and EVs. As their fleet grows, there will be more demand for safety parameters and control systems like the ones developed by MIT researchers.
Amazon has long teased about drone delivery fleets. In the coming year, the company could begin its launch. Notably, the FAA approved the MK30 drones they intend to use for flight last November.
Amazon.com, Inc. (AMZN -0.66%)
Amazon has also been making advancements towards its 10-year goal to deliver packages by drone in the EU. Specifically, the company tested its delivery system in San Salvo, Italy, last December with great results.
If successful, Amazon could use drones to deliver to +500M customers every year. This maneuver would significantly reduce delivery costs and theft, which could result in AMZN stock seeing gains.
Future of MAS Safety Programming
The future of drone safety programming is scalable high-performance AI-based systems. These systems will allow engineers to simplify everything from the creation to the management and distribution of safety protocols.
In the future, your delivery drone will have a full understanding of its real-time environment, flight path, and the location of its co-workers. As such, people continue to approach the day when 24-hour delivery might seem too slow.
Learn about other cool developments in the robotics sector here.
Study Reference:
1. Zhang, S., So, O., Garg, K., & Fan, C. (2025). GCBF+: A neural graph control barrier function framework for distributed safe multi-agent control. IEEE Transactions on Robotics. https://doi.org/10.1109/TRO.2025.3530348