Robocurve

Frontier AI labs are racing to build general-purpose robots within the next few years, yet there is no reliable way to know how capable these systems actually are. The internet is filled with cherry-picked demo videos that do not reflect real-world performance, and no continuous, standardized evaluations for robotics exist. This leaves researchers, forecasters, and the public unable to answer basic questions about where robotics capabilities stand today, how fast they are improving, or when they might start affecting the economy and labor market.

Robocurve runs real-world evaluations of frontier models and robot policies on tasks like bimanual manipulation and refusal of harmful instructions, publishing scores on benchmarks such as RoboDojo-RC Tier 1 and RoboHarm v1. It compares models like GPT-6 Astra, Gemini 4 Argon, Opus 5.5, and GR00T N1.7 across score, cost per attempt, wall-clock time, and release date, including Pareto frontier analysis. The company also builds and open-sources Inspect Robots, a framework (MIT licensed) for running any model on any embodiment on any benchmark with full trace logs and live visualization, with integrations for ROS, Isaac Lab, Cap-X, and XPolicyLab. Robocurve structured itself as a Public Benefit Corporation to report on robotics progress rigorously and neutrally.

Frontier labs are racing to develop general-purpose robots within the next two years, and the page argues this scale of technological change must be measured in the open before it arrives.