Finalist
GAIA
GAIA is an OpenEnv-compatible geospatial reasoning environment where an AI agent infers hidden real-world locations using multi-step tool calls (terrain, weather, sun angle, language, architecture, street/aerial cues) un
Core trick
Agent infers hidden coordinates by spending budgeted tool calls on terrain, sun-angle, and imagery cues, scored by a GRPO reward.
Description
GAIA is an OpenEnv-compatible geospatial reasoning environment where an AI agent infers hidden real-world locations using multi-step tool calls (terrain, weather, sun angle, language, architecture, street/aerial cues) under strict action budgets. It includes an oversight agent that flags contradictions and overconfidence, plus a GRPO training pipeline (TRL + OpenEnv) to improve policy rewards over time, with a live Cesium interface for real-time demo and evaluation.
README excerpt
--- title: GAIA GeoGuess emoji: 🌍 colorFrom: blue colorTo: indigo sdk: docker app_port: 3001 dockerfile: Dockerfile.hf pinned: false --- # GAIA An OpenEnv RL environment for geospatial reasoning under uncertainty — infer hidden real-world coordinates from evidence, with a second agent watching the reasoning. An agent must guess a latitude/longitude by combining terrain, weather, sun angle, language, architecture, and street/aerial imagery — under a hard budget of steps and guesses. [Demo video](https://www.youtube.com/watch?v=inIIm3ZqBIg) · [live demo](https://http--gaia-app--bhzgzs6hgvmy.code.run/) ## How it works - **OpenEnv-compatible** — standard reset/step/state/schema contract plus a WebSocket interface, so the same environment runs local, containerized, and hosted, and policies evaluate reproducibly. - **Actor + oversight, not just a policy** — a separate oversight agent scores each guess and emits structured flags (`LAZY_GUESS`, `CONTRADICTION`, `THIN_REASONING`, `OVERCONFIDENT`), rolled up per episode to `CLEAN` / `CAUTION` / `UNRELIABLE`. Scalable oversight is a first-class runtime output, not a post-hoc metric. - **Multi-tool episodes** — weather, terrain/elevation, sun-angle, language, building-style and map-imagery tools composed in one loop; the agent has to sequence calls and update beliefs before it can guess. - **GRPO training** in `geoguess_env/`, with a Cesium-based real-time visualization and control UI in `worldview/`. - **Live state streamed** over