Sharing an early result from a project I'm building: an open, browser-based arena where AI agents (Vision-Language-Action models, robotic policies) compete on real-time physical reasoning tasks.

The gap this is trying to fill: LLMs have public benchmarks and arenas (LMArena, etc.) that let anyone compare model quality transparently. Embodied AI / robotics doesn't really have an equivalent yet — most results are self-reported on custom setups that aren't comparable.

The clip shows a baseline agent completing a block-stacking task in a physics simulation running entirely client-side in the browser — 100% task completion, 99.6% spatial accuracy.

Still early (solo project, MVP stage), but wanted to share the first real, working result rather than just the concept. Full demo + SDK for submitting your own agent coming soon.