A robot-learning lead deciding whether to hand over data needs one answer: does the simulator rank my policies the way my real robot would? One Robot's homepage shows a single 60-second rollout and an email link, with no number, benchmark or result anywhere. Meanwhile Runway's robotics page states a 0.95 correlation between simulated and real scores across eight policies, from 1,450 rollouts. Visitors can watch a rollout that looks real, but nothing on the page shows that its outcomes match the real robot, and fidelity is the claim the whole product depends on.
First action
Choose the folding task and list the checkpoints you already have from customer or in-house training runs.
How to measure it
Qualified pilot requests from the homepage
- Now
- –
- Target
- 8 a month
- By
- 60 days
Steps
Pick task and checkpoints
One folding task, 6-10 checkpoints ranging from weak to strong.
Co-founder & CTO 1 day
Fix the protocol
Same starting scenes, success definition and number of trials in sim and on the robot.
Co-founder & CTO 1 day
Run the real trials
At least 10 trials per checkpoint on the physical robot; log time spent.
Machine Learning Intern 1 week
Run the simulated trials
At least 50 rollouts per checkpoint in the world model.
Co-founder & CTO 2 days
Compute the card
Rank correlation, success-rate gap, rollouts, robot-hours replaced.
Co-founder & CTO 1 day
Pick honest clips
One predicted failure and one miss, side by side with real video.
Co-founder & CEO 1 day
Publish
Scorecard under the rollout, plus a one-page method note.
Co-founder & CEO 1 day
Use it everywhere
Outreach emails, the pilot request page and the YC profile.
Co-founder & CEO ongoing
Why it works
Pick the task already on the site (box or textile folding). Take 6-10 policy checkpoints of different quality, run each on the real robot and in the world model with the same starting scenes, and publish a small scorecard under "See it in action": rank correlation between simulated and real success, the average gap in success rate, number of rollouts, robot-hours the simulated runs replaced, and two side-by-side clips, one where the simulation predicted a real failure and one where it got it wrong. Link a one-page method note, and use the same card in every outreach email and the pilot request (idea 2).
Technical buyers trust measured numbers over demo videos, and a competitor has already set the expectation that a fidelity number exists. Doing it on deformable, contact-rich folding, which Runway's single-arm tabletop study does not cover, makes the number hard to compare away. It also turns the Founding Robot Learning role's job (validate the world model end to end) into a public asset.
Ready to use
ChecklistSim-to-real scorecard template
Task: box folding (or textile folding) Robot and cameras: <arm, grippers, camera count> Checkpoints compared: <n> Real trials per checkpoint: <n> Simulated rollouts per checkpoint: <n> [ ] Rank correlation, simulated vs real success: <value> [ ] Mean absolute gap in success rate: <value> [ ] Robot-hours the simulated runs replaced: <value> [ ] Clip A: failure the simulation predicted, real vs simulated [ ] Clip B: case the simulation got wrong, and why [ ] Method note: scenes, success definition, date, version [ ] Next update date
Risks and how to handle them
The correlation comes out weaker than hoped
Publish it with the failure analysis anyway, and frame it as version 1 with a dated update plan; an honest number beats none.
Customer data cannot be shown
Use in-house demonstration data for the published card and keep customer runs private.
Evidence
- Your websiteHomepage (checked November 10, 2026): H1 "World models for robots"; "We build simulation environments that are realistic to see and realistic to interact with"; three steps: train a world model on your robot's data, find failures by rolling out your policy, and generate synthetic trajectories that target those failures; "Built for hard manipulation": contact-rich manipulation, deformable objects and precise assembly; "See it in action" shows one 60-second autoregressive rollout; the only call to action is "Let's work together" with a mailto link to the founders' address; no customer logos, results, benchmarks, papers, blog, pricing or form; one "Backed by Y Combinator" badge.Source
- Competitor researchRunway GWM-Robotics (published February 27, 2026): a world model for evaluating robot policies without hardware; Runway reports a 0.95 Pearson correlation between simulated and real-world progress scores across eight VLA policies on a Franka Panda arm (RoboArena tasks), from 1,450 simulated rollouts and over 16,000 human ratings, and notes it measures ranking, not absolute success rates; teams request access through a form.Source
- Y CombinatorLaunch post: the founders found that evaluating a new policy took longer than training it, because of scene setup, trials and resets, and that existing simulators are either not photorealistic or not physics-realistic. The ask is to connect with teams training robots for real-world manipulation, by email. No customers or results are named.Source
- HiringFounding Robot Learning (Ashby, San Francisco on-site, posted August 13, 2026): train manipulation policies (VLAs, diffusion policies, imitation and RL) and run them on real robots, validate the world model end to end and find where sim-to-real transfer fails; "We work with real customer data, real failures, and real deployment pressure." The postings say the company is expanding into policy training.Source