Shepherd Robotics takes on the wiring work behind AI infrastructure

The rush to add computing capacity still depends on someone routing wires, operating equipment controls and checking that everything responds correctly. Shepherd Robotics is building robots to take on that physical work, starting with data centers and electrical manufacturing.

The startup, part of Y Combinator’s Fall 2026 batch, is targeting tasks that require more flexibility than a fixed production line typically allows. Its launch describes robots working under human supervision, with qualified technicians directing the process and handling exceptions. That puts the initial emphasis on helping existing crews get more done.

Two workflows define the starting point. In data center commissioning and inspection, the robots would operate controls, check equipment responses and collect photographic evidence. In electrical assembly, the work includes routing wires and inserting terminals. Both involve repetitive movements, but the parts, layouts and circumstances can change between jobs.

Watch Shepherd Robotics’ original launch demonstration.

The company’s website adds some useful detail about where the technology stands. It shows bench demonstrations involving rack cabling, fuse holders, switches and power distribution units, while labeling other scenarios as simulations. It also explicitly identifies its advertised capacity gains and error reductions as illustrative models or proposed targets, rather than measured customer results.

That distinction matters in a category where a convincing demonstration and a dependable installation are different milestones. A technician needs to know when a robot can complete a task, when it needs assistance and what evidence it leaves behind. Shepherd’s launch does not disclose deployment counts, customer names or measured production performance.

Its technical approach combines robot foundation models with modular hardware. According to its public company profile, Shepherd trains its systems using real-world and simulated data to learn manipulation skills across different workflows. Human interventions are intended to improve the system over time, expanding the work that can be automated.

Shepherd also plans to supply the practical pieces around the models: hardware, tooling, integration into existing workflows and ongoing support. For an operator, that means the proposal covers the equipment and its installation, not simply software that a factory must figure out how to deploy on its own.

The founders bring experience from autonomous driving and robotic manipulation. CEO Feifei Duan previously worked at Waymo and Chef Robotics, including deploying robots in food factories. CTO David Fang worked on robot learning at Dyna Robotics and machine-learning systems at Cruise. Co-founder Kai Ru’s background includes autonomous systems at Cruise and Rivian.

Those backgrounds explain the focus on adapting to messy physical environments. Electrical assembly and commissioning demand careful manipulation, but they also require a system to recognize when the expected sequence no longer matches what is in front of it. Shepherd’s human-supervised approach leaves those exceptions with qualified people.

The team is now seeking conversations with data center, construction and manufacturing operators facing repetitive-work bottlenecks. The immediate test is specific: whether its robots can handle enough of those daily tasks reliably to free technicians for the problems that still need their judgment.

‍