Insider Brief
- Robocurve raised $10 million in seed funding to expand its independent testing of how well frontier AI models control real robots and to support open robotics benchmarks.
- Initialized Capital led the round, with participation from Notable Capital, Decasonic, Y Combinator, Halcyon Futures and others, while Robocurve plans to expand its research team, robot testing and academic benchmarking programs.
- The company said large language models are improving quickly on robotics tasks, citing tests where they outperformed some specialized VLA models and output-token speeds rising about 2.1 times per month, though latency and generalization to more complex robots remain open questions.
Robocurve has raised $10 million in seed funding to expand its work evaluating how well frontier artificial intelligence models control real robots. As a Public Benefit Corporation, Robocurve indicated it operates as a third-party auditor and develops open-source tools and benchmarks designed to measure what AI models can do when connected to physical machines.
Initialized Capital led the round, with participation from Notable Capital, Decasonic, Y Combinator, Halcyon Futures and other investors. The company plans to use the funding to grow its research team, test a wider range of robots and tasks and support academic researchers developing open robotics benchmarks.
Robocurve said its evaluations are designed to remain independent from the AI companies whose models it tests. The company sets its own research agenda and methodology and publishes results without giving frontier AI labs control over the findings.
“The widespread deployment of general-purpose robots could have profound implications for the social contract and the labor market, and deserves a whole-of-society approach to help families prepare for the changes ahead, even if the precise timing of the arrival of such technology still has large uncertainties,” the company noted in a blog post announcing the funding. “When the clouds are heavy, we don’t need to know exactly when it will rain to carry an umbrella.”
AI Models Controlling Robots
As general-purpose AI models increasingly show an ability to perform robotics tasks that previously required models built specifically for robotics, Robocurve said its research has found that large language models can outperform some vision-language-action, or VLA, models on simple robotic tasks.
The company said more recent frontier models have also completed tasks that earlier systems struggled with, though Robocurve noted that major technical questions remain. Among them are whether those capabilities will extend to robots with more complex bodies, such as dexterous hands and humanoids, and whether inference speeds can become fast enough for real-time control.
Robocurve pointed out that latency remains one of the main limits on using large language models to control robots directly.
Robocurve said frontier language models are also improving quickly at controlling robots. The company pointed to tests in which large language models outperformed specialized vision-language-action models on simple tasks and said some tasks that models struggled with a month earlier could be completed easily by newer systems. Robocurve also cited a roughly 2.1-times monthly increase in output-token speed for frontier models, arguing that faster inference could make real-time robot control practical sometime between late 2026 and 2029 if the trend continues.
Robocurve
The company was incorporated about three months ago and said its research has since received more than 6 million views. Its Inspect Robots open-source evaluation framework has been downloaded more than 97,000 times, according to the company.
The company operates a benchmarking program that provides academic groups with funding and robot hardware to develop open-source evaluations. Robocurve said it is making a total of $500,000 available through the program and providing YAM robotic arms to participating research teams.
Robocurve also said researchers from more than 200 institutions have signed up to build benchmarks through its academic program, including participants from 19 of the world’s 20 highest-ranked universities.
