Skydio Flies a Drone Fleet From One Laptop

Bargoti's stated subject is agentic orchestration — using it to command many different drones to carry out the tasks ahead. His method for the next twenty minutes is a live fleet rather than a deck.

On the left side of the screen runs a real-time view, and he begins the demo from it.

The stack Skydio has been building answers three separate questions: how a vehicle operates autonomously, how the cloud system operates, and how the cloud servers work. Bargoti then asks where the different levels of intelligence and automation need to sit for any of it to hold up.

His illustration of the split is a drone responding to something directly in front of it while, at the same moment, someone asks about an incident on the other side of the country. The Colorado launch running during the talk is a simulated inspection after a fire on power lines.

He tells it to track one and narrates it as a possible runaway car they needed to follow.

During the tracking the other drone activity keeps running, and he makes clear he is not touching the controls.

Bargoti divides drone history into three eras. Fifteen years ago the drone was a hobbyist object at home — something to tinker with and to work on at the level of the controlling software.

About ten years ago they became a tool. Industries carried them in the truck, went out and deployed them in the field.

The era Skydio is working toward is drones as infrastructure. For the purposes of the conference he calls these systems agents — physical, embodied, available at all times and anywhere, able to automatically launch, execute and carry out their tasks.

Mid-demo Bargoti tells every drone to pause and return to dock. They all begin heading back.

With the fleet recovering itself, he starts into the presentation itself.

Skydio works closely with SFPD, and Bargoti contrasts the usual response with the drone response. A stolen car normally produces a high-speed chase through the city — dangerous work — whereas a drone can follow the car while the people inside never know it is there.

The footage he shows catches a suspect changing license plates: getting out tools, coming back, tilting the plate up so the drone can read it, then fitting a new one. Throughout, the people in the car are unaware they are being followed. Bargoti's point about the consequences is that behavior in public is very different and operations are a lot safer.

The docks are in use across many industries, and Skydio has started treating them as infrastructure that runs day in and day out, at night, in rain and in sunshine. Some are deployed in Alaska's cold and some in Texas heat.

The reliability figure the company works to is 99.9999%, verified through simulation and testing.

The scale of the deployment matters too: around 16 million people live within a two-mile radius of this infrastructure. For public safety and power companies, the effect is responding to an incident without having to travel there.

The reframing Bargoti describes moves away from first-person-view engagement with a drone. In its place is a multi-agent view that commands the whole fleet against an objective, with the flight itself no longer the operator's concern.

He checks the live demo partway through the talk and confirms the aircraft have landed.

According to Bargoti, that full-stack position is what lets the company decide how its underlying vision system should work in order to see the environment and behave correctly in it.

The conditions that vision system has to survive are named directly: high altitudes, cloud, high speeds and rain.

The footage in the bottom-left corner shows navigation in cities and large-scale planning. GPS in cities, he notes, is poor even on phones, which is what makes robustness there the hard part.

Bargoti calls this the learning flywheel: fly, collect data, operate, come back. Every flight logs information, which he compares in scale to Google Street View.

Before any of it is used there is a sanitizing step to make sure no private information is left in the data, and customers have to know exactly what they are sharing.

The data is dominated by failure cases — where Skydio instructed one thing and the drone did another, or expected something that did not happen. Those cases feed reinforcement learning systems that retrain, evaluate and redeploy.

Skydio is investing in getting the best possible video quality to its servers in areas where bandwidth is thin. The technique is to encode information into smaller sizes and then decode it back into something clear.

The aim is much higher quality under exactly the same network conditions — an investment Bargoti frames as necessary for running a cloud-based infrastructure at scale.

Waymo does local perception, but it also does global planning. The analogy Bargoti uses is travelling from part A to part B of a city — that cannot mean hitting every building, because there is an optimal path.

Prior information is merged with building data and with vector data such as where power lines and roads are. Different behavior may be wanted in different areas.

The drone carries knowledge of this map at all times so it can navigate around within it.

The second model is a more conventional machine learning inference model. Its job is to understand objects in a scene and track them through occlusion.

The implicit representation involved is that an object gone behind a building may come out on the other side. The behavioral consequence is that the drone navigates to follow it there, or moves in a different direction to do so.

Bargoti notes that roughly five years ago this was done in a more conventional, manual way.

Skydio is exploring reinforcement learning techniques and learned end-to-end approaches. Bargoti frames these as a way to avoid having to engineer every edge case that can come up.

The robustness targets he names are rain, snow, day and night — using vision only as the sensor modality. He gives no figures for performance or reliability in this passage.

Much of Skydio's semantic reasoning work targets infrastructure customers, and the question it answers is what is present in the scene. Utility poles are the illustration.

The workflow Bargoti describes: an instruction arrives — go look at this line, something has gone wrong — and the drone has to reach the location, understand the scene and then act inside it.

The team builds primitives and tools that are coded today but that an agent can access at any time, so it can better understand what the drone is seeing and what it might do about it.

The self-driving-style dream — feed raw sensor data and get the right actuation, pointing and path out — is the one Bargoti ties to what the self-driving community and other robotic systems are chasing.

What that output is could take several forms: the actuation that happens, where the drone is pointing, or where it goes. Skydio is testing reinforcement learning with multiple instantiations of the behavior.

Whether it reaches the right end of the spectrum is, in his framing, still open.

The aircraft would be launchable and recoverable from anywhere, with the cloud as the place to build orchestration quickly.

The design gives each system basic APIs of interactivity. Cloud agents come in, tap into those APIs and make decisions on top of them.

The stated end goal is allowing very high-level thinking when working with these drones.