Robocurve

Robocurve, Inc. is a San Francisco public benefit corporation that runs independent real-world evaluations of physical AI. It tests AI models, including large language models (LLMs), vision-language models (VLMs) and vision-language-action (VLA) policies, while they control physical robot arms, and it publishes the results. Its purpose is to measure and report frontier robotics capabilities to the public. As a public benefit corporation, it has a legal duty to evaluate these capabilities independently. The company says the robotics industry lacks well-run, standardized benchmarks.

Its main open-source product is Inspect Robots, an MIT-licensed evaluation framework. The framework lets any LLM or VLA be run on a range of robotic arms against benchmarks, either on real hardware or in simulation. The company reports more than 97,000 PyPI installs. Hardware adapters are published for I2RT YAM bimanual arms (driven by MolmoAct2), LeRobot SO-ARM (SO-101) follower arms (driven by LeRobot policies), and Franka FR3 and Panda arms (driven by openpi DROID policies).

Other open-source projects on its GitHub are KitchenBench, a bimanual kitchen-manipulation benchmark for VLA models; WorldEvals, a curated catalog of VLA and physical-AI benchmarks that can run through Inspect Robots; WorldPolicies, a registry of published robot-policy checkpoints, indexed with evaluation metrics and validated through continuous integration; and ClapboardBench, real-robot evaluation traces of bimanual YAM arms operating a film clapperboard under frontier VLM control.

Robocurve's published evaluations have tested frontier models such as Opus 5, GPT-6 Astra and MolmoAct2 on bimanual manipulation tasks. The tasks include Tower of Hanoi, block stacking and Jenga. The company says its research drew more than 6 million views in three months and that more than 200 institutions, including 19 of the world's top 20 universities, have taken part in building its benchmarks. Its website lists MIT, Stanford, Harvard, Princeton, Caltech, the University of Washington, MATS, Foresight Institute and the NVIDIA Inception Program among its partners and supporters.

Robocurve was founded in 2026 by CEO Jay Chooi and is in Y Combinator's Summer 2026 batch. The team has three people. Chooi is a Rhodes Scholar with an MA in Statistics and a BA in Computer Science and Mathematics from Harvard. He was previously a researcher at the UK AI Security Institute and a top contributor to Inspect Evals, and he has published at ICML, EMNLP, ACL and ACM EC.

On September 14, 2026, Robocurve announced a $10M seed round. Initialized Capital led the round, with participation from Notable Capital, Decasonic, Y Combinator, Halcyon Futures and others. The company plans to use the funds to expand its research team, test more robots and tasks, and support the development of academic benchmarks. That support includes a total of $500K in funding and free YAM arms for academic groups.