Skydio

The opening demonstration moves between docked drones in San Mateo and Colorado, hands-off tracking of a passing car, and a third aircraft at headquarters. Bargoti commands the fleet to return to dock and later reports that all landed. Running over conference Wi-Fi makes the central requirement concrete: safe execution must continue while the operator attends elsewhere, even if the laptop closes. He argues that assigning a skilled pilot to every aircraft creates a staffing bottleneck, while acknowledging that individually launching several drones still imposes cognitive load. Objective-based, potentially unsupervised orchestration is presented as a direction beyond this demonstration.

Bargoti describes thousands of docked systems deployed with utilities, public safety organizations, and construction companies, with about 16 million people living within two miles of the infrastructure. His examples include a utility patrol finding a pole burning internally and SFPD observing a reportedly stolen car while its plate is changed, allowing officers to position themselves before intervening. These are speaker-reported examples, without supplied measurements of inspection effectiveness or pursuit safety. Docks in Alaska and Texas illustrate environmental demands. The stated 99.9999% reliability requirement has no defined event or denominator here and should not be read as a measured fleet success rate.

Immediate autonomous actions run on the aircraft; cloud servers receiving video and telemetry can support heavier inference and longer-term planning. That division makes bandwidth consequential: Bargoti describes encoding smaller video streams and decoding clearer imagery under constrained network conditions, without specifying the method. Minimal tracking remains on-device, while heavier cloud models can provide broad movement guidance with one-to-two-second feedback latency. Flight logs also expose differences between intended and actual behavior, feeding retraining and evaluation. He identifies removal of private information and clarity about customer data sharing as requirements for this learning cycle, rather than supplying evidence of their implementation.

Maps combine prior building information with road and power-line data to support global planning alongside local perception. A construction-site example shows why maps need maintenance: the recorded scene no longer matches prior knowledge. Returned flight observations feed synchronization so subsequent fleet operations can use an updated map. Object tracking addresses a different uncertainty: when a target disappears behind a building, its estimated hidden state and likely emergence point guide the drone’s repositioning. Bargoti discusses learned approaches to these movements while retaining the challenge of vision-based operation across weather and lighting conditions.

In the white Jeep example, a vision-language model uses a drone API to command a search trajectory, examines the resulting scene, and invokes tracking and following tools after finding a candidate. Utility inspection similarly depends on primitives that help identify scene elements and choose actions around them. The agent composes exposed capabilities rather than requiring a bespoke branching strategy for each request; those tools still require implementation. The recording does not specify the production deployment scope of every VLM or orchestration capability.

Bargoti describes reinforcement-learning exploration toward direct sensor-to-action behavior, but says fully end-to-end systems remain difficult to diagnose and give reliability guarantees for physical operation. Search-and-rescue branches—inspect trees, look underneath, enable thermal sensing, search elsewhere—illustrate where agents using high-level tools might reduce manually written strategies. He closes by extending the infrastructure approach to smaller quadcopters and exploration of fixed-wing aircraft, with interaction APIs enabling cloud orchestration, then gives a recruiting invitation across hardware, software, autonomy, and networking. These closing remarks describe the recording’s direction and invitation, not verified current hiring or deployment status.