Avea Robotics and the 10-millisecond human keeping robot fleets alive

A robot on a warehouse floor is a wonderful thing right up until it isn't. It picks, it sorts, it hums along on its own - and then it meets a pallet wrapped the wrong way, or a box the training data never showed it, and it simply stops. On a busy floor, a stopped robot is not a pause. It is a clock, and every minute on that clock costs money.

Avea Robotics was built around that exact moment of failure. The San Francisco company, part of Y Combinator's Spring 2026 (P26) batch, makes software called Sentinel that lets a human operator drop into a stuck robot from anywhere on earth and drive it back to work. The pitch is unglamorous and, once you hear it, hard to argue with: autonomy is close, but it is not finished, and the gap between "mostly works" and "runs a business" is where robots go to die.

The ProblemThe last five percent is the expensive part

Robotics companies have gotten good at demos. The trouble starts when a robot leaves the controlled world of the lab and enters a warehouse that was designed for people, not machines. Edge cases pile up. A robot that is right ninety-five percent of the time still needs a plan for the other five, and until recently that plan had a name: hire someone to stand near it.

Keeping staff on standby to babysit machines quietly destroys the economics of a deployment. You bought a robot to remove a cost, and now you are paying a person to watch it. Avea's founders describe the failure plainly.

"Robots aren't perfect - they make mistakes and get stuck in edge cases. That means lost time, and on a factory floor or inside a logistics warehouse, that can translate to thousands of lost dollars." Avea Robotics, launch post

The ProductSentinel, and the tyranny of lag

Sentinel is teleoperation software, but the word undersells the engineering problem. Anyone who has tried to control something over the internet knows the enemy is lag. A quarter-second delay is fine for a video call and useless for a robotic arm reaching for a fragile part. Miss the timing and you have knocked the thing over instead of saving it.

So Avea's whole game is latency. The company says Sentinel can run as low as ten milliseconds - fast enough that an operator reacts to the robot's world in something close to real time. Around that core it layers the things a remote driver actually needs: immersive depth perception, haptic feedback so you can feel resistance you cannot see, and up to six full-HD video feeds so nothing important happens off-camera.

Getting a robot onto Sentinel is deliberately boring, which is a compliment. A customer hands over the robot's URDF - the file that describes its joints and geometry - plus a joint-control API. Avea deploys the software as a Docker container. Then it goes live. The company runs a demo of office robots you can drive yourself before you commit.

And it is not fussy about hardware. Sentinel drives standard six-degree-of-freedom arms, heavier industrial arms, seven-degree-of-freedom setups with full-arm retargeting, and humanoids - two hands, a torso, and locomotion all at once.

The BusinessDowntime, quietly turned into a dataset

There is a second idea tucked inside the first, and it is the more interesting one. Every time an operator takes over a stuck robot, Avea captures what a skilled human did in a situation the AI could not handle. That is not just a rescue. It is a labeled example of exactly the behavior the next model needs to learn.

"Robotic AI models will never get better if they just sit in the lab. Deployments are the bottleneck for autonomous robots." Avea Robotics

The model is business-to-business software. Robotics and Physical AI companies pay recurring fees to keep fleets running and to collect that intervention data, and the revenue grows with the number of robots under management. At launch Avea reported roughly eighteen thousand dollars in monthly recurring revenue and said Sentinel had helped deploy nearly thirty robots - small numbers, but paying ones, and the first deal only closed in March 2026.

The FoundersFrom satellites to the shop floor

Avea is the work of two University of Illinois engineers, Ary Indarapu and Vikram Vadrevu, who left Urbana-Champaign for San Francisco with a fairly specific grudge against labs. Indarapu, a computer engineering graduate, is chief executive. Vadrevu is the one whose resume makes the latency obsession make sense: before robots, he wrote flight software for NASA and Department of Defense satellites, a field where a few milliseconds is never an abstraction.

The company did not start with a Y Combinator check. It started as a student team that placed sixth and won fifteen thousand dollars at the University of Illinois College New Venture Challenge, one of two Grainger Engineering teams to secure funding that round. A year later it was in the P26 batch with Harj Taggar as its partner.

The MarketThe human layer everyone forgot to build

Most of the noise in robotics is about two things: better models and better hardware. Avea is betting on the unfashionable third thing - the human layer that sits between them and keeps the whole system honest while the models catch up. The alternative to Sentinel is not a competitor so much as a decision: a robotics company builds its own remote-intervention stack and staffs its own standby crew. Avea's argument is that this is the plumbing, not the product, and you should buy it rather than build it.

Whether that argument holds depends on how long the "mostly autonomous" era lasts. Avea is, in a sense, selling insurance against robots not being ready yet - and quietly using the payouts to help make them ready. If autonomy arrives slowly, that is a long and useful runway. If it arrives fast, the intervention data Avea collected on the way there is not a bad place to be standing either.

"Our software is ultra low-latency, meaning operators can react to the environment in near real-time." Avea Robotics

For now, the company is doing the plain work: getting robots out of the lab, into warehouses and factories, and keeping a person one click away when they get stuck. It is not the flashiest corner of the robot boom. It may be one of the more necessary ones.