Serve Robotics Opens Fleet Software and Data to Physical AI Companies

Serve Robotics announced Physical AI Services on November 5, 2026, making its robot-connectivity, data-curation and real-world training-data tools available to other companies building physical AI. Bedrock Robotics, which develops autonomous systems for heavy construction equipment, is among the launch customers.

The three-product suite productizes infrastructure Serve built while expanding its own operations. The company says it has deployed more than 2,000 robots in more than 40 cities and 27 hospitals, and that Physical AI Services already powers thousands of vehicles across a dozen industries. Those include construction, mining, autonomous trucking and Serve’s sidewalk food-delivery fleet, giving the offering an operating base that spans different machines and environments.

“Building a robot is only the beginning. Keeping thousands of them connected, learning and improving requires another layer of technology,” said co-founder and CEO Ali Kashani.

Connectivity and Fleet Supervision

Serve Link connects machines in the field to models, monitoring and training systems as more physical AI computation moves into data centers. Its functions include task management, unit-status monitoring, remote assistance for exception handling, data collection and hard-example mining. Bedrock uses Link to connect equipment over Wi-Fi, LTE, 5G and satellite networks.

The Physical AI Services documentation says the software can bond 4G, 5G, Wi-Fi, satellite and private cellular connections, predicts optimal transmission logic to avoid buffering, and is designed to reduce latency, processing time and compute overhead while running on almost any hardware. Those capabilities are intended to keep fleets reachable across varying locations and network conditions.

From Fleet Data to Training Data

Serve Data Curator addresses the volume of information produced by deployed robots. It converts petabytes of sensor input into searchable, visualizable fleet data, allowing engineers to find relevant scenarios, investigate failures and identify long-tail edge cases without manually reviewing every recording.

Selected examples can be assembled into focused datasets for training and evaluation. Integrated annotation and connections to machine-learning workflows link the discovery of useful examples to their use in model development. Model performance can then guide which examples teams curate, collect and simulate next, connecting what robots encounter in the field with subsequent training.

Serve says Data Catalog contains millions of hours of real-world data from robots operating on streets, including topographical and infrastructure information, and that live-deployment collection respects local laws and the privacy of people nearby. Terrain and infrastructure data from sidewalk deployments are intended to teach foundation models about complex human environments such as city sidewalks.

The suite is modular and plug-and-play, with documented uses in humanoid robotics, agriculture, forestry, underwater exploration and autonomous trucks alongside delivery and construction. Its common thread is infrastructure: connectivity, data transport and curation are offered as existing foundations so customers can devote engineering time to their own machines and autonomy.

For Serve, the central question is how far that operating experience can transfer. The offering asks customers to build on infrastructure and data shaped by sidewalk and hospital deployments while their own machines tackle different tasks; its value rests on how much engineering work those shared foundations can remove.