Real-world evaluations of physical intelligence
Robocurve is a Public Benefit Corporation helping society understand the state of robot intelligence. We build open-source tools and independent benchmarks to evaluate how capable AI models and embodiments are in the real, physical world.
Why it matters
Frontier labs are targeting general-purpose robotics by 2028, yet the field of robotics evals barely exists. Labs evaluate in-house, so no one actually knows how good anyone else is, or where the real frontier sits.
The arrival of general-purpose robots could drastically transform society. We want the public’s understanding of that future to be grounded in rigorous, verifiable data that is independent of any agenda.
Our work
We build open-source tools and benchmarks for evaluating robots.
Inspect Robots
Run any model on any embodiment on any benchmark, with full trace logs and live Rerun visualization.
If you know Inspect AI, this is that for robotics.
- Real-world first, with simulation support
- Supports running LLMs and coding agents such as Claude Fable 5 to control robots
- First-class integration with ROS, Isaac Lab, Robolab, Cap-X, and XPolicyLab's 40+ VLAs
- Open-source license (MIT)
World Evals
A catalog of benchmarks in the physical world, from making a sandwich to building a data center.
If you know Inspect Evals, this is that for robotics.
- Compatible with any form factor: arms, dexterous hands, humanoids, AMRs, quadrupeds, etc.
- Real-world first, with digital twins for simulation runs
- Benchmarks have well-specified distributions and reproducible setups
- Open-source license (MIT)
KitchenBench
Evaluating robots on ten bimanual kitchen tasks, including pouring, lid removal, folding, part-mating, a two-arm handover, and tool-mediated scooping. Includes a dependency-free mock kitchen.
GitHubDataCenterBench
Evaluating robots on data-center construction tasks, including racking servers, routing cable, and the physical work behind the compute supply chain. Informed by the Texas Electrician Exam.
Contact usLaundryBench
Evaluating robots on ten laundry tasks centered on deformable manipulation, including folding t-shirts, hoodies, and pants, plus sock pairing and sorting lights from darks.
Contact usCoffeeBench
Evaluating robots on coffee-making tasks, from grinding and tamping to pulling a shot and steaming milk.
Contact usWe welcome open-source contributions.
Contribute on GitHubAbout
Robocurve is a Public Benefit Corporation helping society understand the frontier of physical AI.
Our mission
Our mission is to independently evaluate and transparently report the real-world capabilities of robots and the AI models that control them.
Our team
Our team brings experience from AWS and the UK AI Security Institute, with published research at ICML, ACL, EMNLP, and ACM EC.
