From Rollout to Reset: A Graph-Based Harness for Autonomous Long-Horizon Manipulation Evaluation
HALTER scores a long-horizon manipulation rollout on real hardware and restores the scene for the next one without a human. It builds a spatial scene graph online from RGB-D keyframes, and an LLM reasons over that graph to score the rollout, to plan a reset over a library of learned atomic reset skills, and to verify that the reset succeeded.
Project page: https://intuitive-robots.github.io/HALTER/
main— the codebase.gh-pages— the project page. Its source lives inwebsite/, and the built site sits at the branch root, which is what GitHub Pages serves.
Code release is in progress.