Enact · robotics lab
10×Fewer evaluations for the same answer.
Their earlier process took roughly 300 to 400 evaluations to get useful engineering feedback on a model change. With verified starting conditions, about 30.
Physical AI can't
improve what it
can't measure.
Robots now learn like software. They still get tested like it's 1995: by hand, by eye, one trial at a time. Prism is the measurement layer for physical AI. Every robot that ships will be tested on it.
The shift
Every robotics lab is becoming a model company. They train, change a checkpoint, and train again. Software teams ship that loop a hundred times a day because they have a test suite. The physical world has no test suite. So labs rerun trials, argue about results, and ship on a feeling.
$10B+
Invested in robot foundation models since 2024. Every dollar of it needs a way to measure progress.
100+
Labs training generalist robot policies today, from startups to the largest AI companies.
~40%
Of a robot learning team's time spent running and resetting physical evaluations.
0
Standard ways to test a robot policy in the real world. Prism is building the first.
The missing control
A few centimeters
change the answer.
Did the model get better, or did someone put the cup in an easier spot? When the setup moves between runs, a real improvement disappears into the noise. The only fix today is brute force: hundreds more trials.
Same two models. Model A scores 60, model B scores 70.
The results overlap. Nobody can say B is better without hundreds more trials.
Illustration
50%+
Swing in success rate from centimeter-scale shifts in where objects start.
~30%
Disagreement between human scorers watching the same trial.
1 in 5
Episodes thrown out because the scene was set up wrong.
10×
More trials needed to see a real policy change through the noise.
What Prism does
The test starts
before the robot moves.
Test code can name a position. It can't make anyone put the cup there. Prism closes that gap: it projects the setup onto the table, checks it with cameras, and ties every result to the exact scene that produced it.
Define the case.
Placements come from the lab's own code through our SDK, or Prism generates a set for the task.
Project, place, verify.
The setup appears on the table. Cameras confirm identity, pose and tolerance before anything runs.
Run the robot.
Any arm, any policy. Every model faces the same verified starting conditions.
Keep the evidence.
Outcome, model version and verified setup are stored together, so a result can be trusted, compared and replayed.
Proof
Fewer trials.
A clearer answer.
Prism is paid for and running at Enact, New Theory and Trossen Robotics today: 10 paid units across 3 paying customers. Every customer that installed it kept it, and two of them measured exactly what changed.
New Theory · 30-day measurement
40% → 3%Evaluation variability.
Same test, run again and again. Consistent physical starting conditions made it possible to see whether the next model actually improved.
10Paid units installed
3Paying customers
20Stations running in labs
The data opportunity
Find a failure once.
Use it forever.
A single success rate hides everything useful. Because Prism verifies every starting condition, each trial becomes a case a lab can run again: on another station, after the next update, in another building. That library of real-world cases is the asset, and it compounds with every run. Over time it becomes the benchmark the whole industry reports against: the ImageNet of physical AI.
Illustrative workflow
Selected physical condition
Position C4
x 30 cm · y 40 cm
Earlier50%10 / 20 valid trials
New85%17 / 20 valid trials
+35 points at the same starting position.
The average went up. Look at the map and two positions got worse, and eight were never tested. That is what a single number hides.
The business
A station in the lab.
Software on every run.
Hardware opens the door. Every evaluation after that runs through Prism's software, and a lab re-tests every time its model changes, so revenue recurs with its training cadence. Enact started with one station and runs five today.
Prism Station
$1,500 per workcell at an 80% margin. Mounted over an existing workcell, it works with the robots a lab already has.
Software on every run
Every staged, verified and scored evaluation is billed, and every engineer on the team works in the console, replay and case library.
Tooling and enterprise
The SDK and scoring API reach robots outside a station. Multi-site labs sign enterprise contracts for fleets, single sign-on and on-prem scoring.
Companies building robot foundation models.
16,000 humanoids were installed in 2025, and 100,000 are expected by 2027. CB Insights ↗ Counterpoint ↗
Robots a data team runs at once.
Each workcell under test is a station. Google's robot fleet ran 23 robots through 4,800 evaluation trials. Trossen Robotics ↗ Google Research ↗
Enact's stations after the first.
Labs buy again once Prism is in the room. Every customer that installed it kept it.
Applied Intuition, the same play for cars.
Testing and validation software for self-driving cars, valued at $15B. Prism is that layer for robots. Bloomberg ↗
Targets
$1 per evaluation at 100 a day, 250 working days. 4 seats a station at $50 a month. $50K a lab for tooling. $150K a lab for enterprise from 10 stations.
Annual software revenue$187M
- Evaluations
- $125M
- Platform
- $12M
- Tooling
- $12.5M
- Enterprise
- $37.5M
The bigger picture
Intelligence that
reaches the physical world.
Under Prism is one capability: a system that sees a surface, understands what is on it, and puts precise light exactly where it is needed. Robot evaluation is where that capability is worth the most today. The same stack already guides people through real tasks, from lab setups to piano lessons, with nothing to wear. Robotics is the wedge. The company is the operating system for every physical workspace: labs, factories, kitchens, classrooms, homes.
Built beside the people
who use it.
“I helped ship Apple Vision Pro, leading two stability labs with about 150 robots and fixtures. I know what it costs when you can't trust a test. I built Prism with the labs that needed it, working next to their operators. Now we're turning those first deployments into the standard way physical AI is measured.”
Built by
Backed by
Founders, Inc.The round
Raising our seed
to own the category.
The product works, customers pay for it, and the labs that use it don't go back. This round turns twenty stations into a fleet, and a fleet into the standard.
100 stations in labs
Productized Prism Station, built and installed for every major robot learning lab in the US.
Self-serve software
Console, scoring and SDK that a lab can turn on without us in the room.
The public benchmark
The first shared library of verified physical test cases, published with partner labs.
Physical AI should iterate
at the speed of code.
Change the model. Recreate the conditions. Measure the improvement. Know what to train next.