Stage
Light shows the setup.
Every object's exact pose is projected onto the table. Operators set a scene in seconds, with no rulers and no tape.
Built by
Robot evaluation,
from setup to results.
Stage and verify the scene, run your policy, and review the results. Prism connects physical test conditions, recordings and scores in one evaluation workspace.
Backed by
Founders, Inc.
Founders, Inc.The full evaluation workflow.
One platform.
Define tasks and scoring rubrics. Assign runs to stations. Stage objects with projected guidance, verify the setup and keep video, telemetry and scores with each episode.
Compare policies, inspect failures and replay saved conditions. Scene history keeps changes to the setup and model traceable across evaluations.
Most robot evals measure the room. A cup a few centimeters off, a different scorer, a missed reset, and the number moves. Teams lose weeks rerunning trials they still can't trust.
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.
Watch an evaluation,
start to finish.
From an empty table to a scored, comparable result. One station, one operator, no guesswork.
One closed loop,
from setup to answer.
Every trial runs the same way, every time. Light shows the setup, vision checks it, a model scores it, and the console keeps it.
Verify
Vision checks the scene.
Identity, pose, occlusion and tolerance are confirmed before the robot moves. Mismatches can be reviewed before a trial is accepted.
Run
Your robot, your policy.
Prism tracks the scene live while the policy runs, so the record of every trial matches what really happened.
Score
Every trial, scored with a reason.
Define scoring conditions for each task step. Review outcomes against the same rubric and keep the score with its evidence.
Compare
Answers your team can ship on.
Results land in the console. Compare policies head to head, catch regressions, and see exactly where a robot fails.
Every run, every position,
one view.
See where each policy wins and fails across the table, how it compares to the last checkpoint, and which runs need a look.
Pick and place · v3 suite Policy B vs Policy A · last 7 days
Pass rate82%▲ 11 pts vs A
Trials scored1,284▲ 312 this week
Stations online20/20All verified
Regressions3▼ needs review
Pass rate by start position50 cells
81839554142523025162557616773847957482650646796899398866934425967677872835950375313053495740303216
FailsMixedStrong
By conditionA vs B
Baseline
Rotated 45°
Clutter
Low light
New object
Policy APolicy B
Recent runsNewest first
| Run | Case | Result | Station | Scored by |
|---|---|---|---|---|
| run-4f2a | cup-to-plate · rot-45 | 9 / 10 | lab-a / 07 | Prism VLM |
| run-4f29 | cup-to-plate · clutter | 6 / 10 | lab-a / 03 | Prism VLM |
| run-4f28 | drawer-open · low-light | 4 / 10 | lab-b / 12 | Prism VLM + review |
Real physics.
Lab-grade control.
Simulation is fast but it is not the world your robot ships into. Hand-run evals are real but they drift. Prism gives you the real world with the repeatability of a simulator.
Manual evalsSimulation & world modelsPrism
Real objects, real physicsYesApproximatedYes
Same start state every trialBy eyeYesYes, verified
Consistent scoringVaries by personScriptedRubric + reason
Setup time per trialMinutesNoneSeconds
Sim-to-real gapNoneAlways thereNone
Results your team can auditSpreadsheetsLogsVideo, pose and score per trial
Faster answers.
Numbers you can trust.
Less noise between runs means 30 controlled trials tell you what 300 brute-force trials used to.
10×
Faster evaluation cycles
The same scene is staged and verified before every run, so each trial counts.
3%
Physical test variability
Down from 40% between repeated runs of the same test condition.
20khr
Engineering time saved
Resets, reruns and result cleanup your engineers never had to do.
“Before Prism we argued about whether a new checkpoint was better. Now we stage the same fifty conditions, run them overnight, and the console tells us. It changed how fast we ship.”
The whole platform,
when you want it.
Open any part to see how it works.
The platformStation, test cases, scoring, console, API and fleet tools, in one system.
Prism Station
Projector, depth cameras and on-board compute in one unit that mounts over any workcell.
- Self-calibrates and locks to the table
- Guides operators step by step
- Runs on site, no cloud required to test
Cases & versions
Define a test once: object poses, task, rubric and success criteria. Replay it exactly, forever.
- A seed and version history for the state of a table
- Sweep one condition at a time
- Share suites across labs
Automatic scoring
A vision-language model scores every trial against your rubric, with a written reason you can audit.
- No scorer drift between shifts
- Human review and override
- Every call stored with its evidence
Light Console
One place for every station, run and result across your whole organization.
- Live fleet and run history
- Pass rates and heatmaps per position
- Policy comparisons and regression alerts
API & SDK
Plug Prism into the stack you already have. Stream runs in, trigger scenes, pull results out.
- Documented REST API with OpenAPI spec
- Run ingest from your own harness
- Embeddable result views
Fleet & security
Built for teams running many stations across many sites, with your data kept private to you.
- Signed over-the-air updates
- Calibration backups and health checks
- Roles, scoped keys, encryption at rest
Capability mapsChange one condition at a time and see exactly where a policy holds up and where it breaks.
Prism changes one condition at a time: position, orientation, clutter, objects or light. You get a map of exactly where each policy holds up and where it breaks, before your customers find out.
Built for every teamWhat research, eval operations and leadership each get from Prism.
Know a checkpoint is better.
Compare policies under identical physical conditions and see real gains, without arguing over whose run was cleaner.
Run more, reset less.
Guided setup and automatic scoring let a small team run hundreds of trustworthy trials a day across every station.
One number per release.
A clear, repeatable score for every model you ship, with the evidence behind it when investors and customers ask.
DeploymentMount, claim, calibrate, run. Scored trials the same afternoon.
Mount it.
Place the station over your workcell and plug it in.
Claim it.
Add it to your organization in the console with a one-time code.
Calibrate.
Prism maps the table and locks the calibration.
Run.
Load your cases and start scoring trials the same afternoon.
In the labReal stations, real operators, real runs.
Prove it on your robot
in 30 days.
We install a station on your workcell, build your first test suite with you, and hand you a report on your own policies. If it doesn't change how you evaluate, you don't continue.
- Prism StationInstalled and calibrated on your workcell
- Your first test suiteBuilt with your team around your real tasks
- Light ConsoleSeats for your whole team, API keys included
- Automatic scoringTuned to your rubric, with human review
- Engineer on callDirect line to the people who built it
- Running in a dayMount, claim, calibrate, run. Scored trials the same afternoon
- Results reportCapability map and policy comparison at day 30
Built by Apple engineers
and Carnegie Mellon researchers.
We have shipped products to millions and run robots in research labs. Now we build the tool we always wanted: a way to set up the physical world exactly, every time, and trust the number that comes out. Based in San Francisco.
Built by
Backed by
Good to know.
Does Prism work with our robot?
Yes. Prism controls and checks the physical scene around the robot, so it works with any arm, gripper, policy or training framework. Results come into the console from the station, or from your own harness through the API.
How is a trial scored?
Prism scores every run automatically. Its vision-language model watches the trial, judges it against your rubric and records the reason with the recording. Your team sets the task steps, scoring conditions and success threshold.
What does it cost?
Every engagement starts with a 30-day pilot on your own robot. After that, stations are priced per workcell with the console, scoring and support included. Book a call and we will size it to your lab.
How fast can we start?
Most teams go from unboxing to scored trials in a day. We help set up your first test suite and stay on call while your team gets going.
Make every eval count.
Policy evaluation, regression testing and capability mapping on one platform.
