Instance Wants to Grade Every Robot - So Nobody Has to Watch

Somewhere in a robotics lab right now, a person is watching a robot arm try to pick up a mug. It reaches, closes, lifts - and the human types "success" into a spreadsheet. Then they reset the mug, reset the arm, and do it again. And again. Thousands of times a week. This is the quiet, unglamorous machinery behind every impressive robot demo you have ever seen, and it is exactly the job Instance is trying to make disappear.

Instance is a San Francisco company in Y Combinator's Summer 2026 batch. Its pitch fits on one line - automated evals for robot policies - but the problem underneath it is bigger than it sounds. To teach a robot anything, you have to know whether the robot actually did the thing. Not almost did it. Not looked like it did it. Actually finished the task. Getting that answer, at scale, is slow, expensive, and boring, which is why almost no one wants to build for it and why it quietly slows everyone down.

Instance's first product is a success detector. You hand it a task description and a video of the robot's rollout. It hands back a verdict: did the robot succeed, and if not, which part broke. It reads the run subtask by subtask, so instead of a single thumbs-up or thumbs-down you get a map of where things went right and where they fell apart.

01 / THE LOOPThe job nobody wants to do by hand

Training a robot policy is a loop, and the loop has a human wedged in the middle of it. The robot rolls out. Someone judges whether it worked. Someone resets the scene. The next rollout starts. Multiply that by the thousands of attempts it takes to move a policy forward, and you get teams of people spending their days as human stopwatches and referees.

🤖

Roll out

Robot attempts the task

✅

Judge

Success detector calls the outcome

↩️

Reset

Scene is returned to start

🔁

Repeat

Next rollout begins

Instance's long-term vision is to close that loop entirely: the robot rolls out, the detector judges success, a second robot resets the scene, and the next rollout starts with no one in the room. Today the company is starting where the pain is sharpest and the work is most repetitive - the judging step.

02 / THE DETECTORBetter than a frontier model at one narrow thing

The obvious way to grade a robot video today is to point a general-purpose vision-language model at it and ask. It works, sort of - but big general models are slow and costly when you are running them across millions of episodes, and they were never built for this specific question. Instance built a detector aimed only at robot evaluation, and the company says it beats frontier VLMs on both accuracy and latency across more than 10,000 human-labeled episodes, seven robot platforms, and eight benchmarks.

Instance

accuracy ▲ latency ▼

Frontier VLM

slower · pricier

Illustrative comparison. Instance reports higher accuracy and lower latency than general frontier VLMs on its internal 10k-episode benchmark. Exact figures not publicly disclosed.

The narrowness is the point. A model that only has to answer "did this robot finish this task" can be tuned to do that one thing cheaply and quickly, which matters enormously when the answer is needed thousands of times a week rather than once. The subtask breakdown is the other half of the value: knowing a run failed is useful, but knowing it failed at the grasp rather than the placement is what actually tells an engineer what to fix.

03 / GROUND TRUTHWhy "ground truth" is the whole bet

Instance's tagline is "Ground truth for robot learning," and that phrase is doing more work than a tagline usually does. Robot policies are trained on episodes. If the labels on those episodes are wrong - if a run marked "success" was actually a near-miss - the policy learns from a lie. As robotics teams scale to millions of episodes, the value quietly shifts from whoever builds the model to whoever can certify which episodes were real. Instance is betting on the certifier.

That framing also explains the product beyond the detector: a verification console where teams submit episodes with task specifications and get back trusted verdicts backed by evidence, turning raw footage into verified training data. It works across any robot and any camera angle, which matters in a field where no two labs have the same rig.

04 / THE FOUNDERSFrom planetary atmospheres to robot report cards

Instance was founded by Claire Mao and Lucy Cai, who have been friends since middle school and studied together at MIT. Mao, the CEO, studied math and computer science; before Instance she built software to simulate planetary atmospheres at NASA's Jet Propulsion Laboratory and researched sustainable propulsion for spaceflight at the MIT Media Lab. Cai, the CTO, holds a master's from MIT's Learning and Intelligent Systems Lab and has worked on satellite software at SpaceX, LLM testing at Amazon, and brain-computer interfaces at Blackrock Neurotech.

The résumés matter less as name-drops and more as a hint about temperament. These are people who have spent years around physical systems that are unforgiving about whether something actually worked - a satellite either phones home or it doesn't. That instinct, applied to a lab full of robot arms, turns into a fairly stubborn question: how do you know, without a human, that the robot did the task? The answer is the company.

05 / THE MARKETPicks and shovels for the robot rush

Robotics is having a moment, and most of the attention - and most of the money - is going to the robot brains: the policies, the foundation models, the flashy manipulation demos. Instance is positioned one layer down, in the infrastructure that makes all of that measurable. It is a picks-and-shovels bet. The company is not trying to build the best robot; it is trying to be the thing every robot team relies on to know whether their robot is getting better.

There is company in that layer - research efforts like AutoEval on autonomous real-world policy evaluation, and other groups working on robot evaluation - plus the ever-present option of teams just wiring up a general model themselves. Instance's wager is that a focused, faster, more accurate detector, plus a verification workflow built for the messiness of real labs, beats the do-it-yourself route once you are running evaluations at serious volume.

For now the company is small and early - a founding team, a launched detector, a public demo, and a spot in YC's Summer 2026 batch. You can try it yourself at the demo: drop in a robot video, or record one from your phone, and watch Instance return a verdict live. It is a modest thing to click through. It is also, if the founders are right about where robotics is headed, a preview of the boring, essential layer the next generation of robots will quietly stand on.