Preprint improves robot navigation model while removing runtime planner

Researchers report that a training method raised an 8B navigation model's success rate from 35.0% to 44.9% across all 3,189 REVERIE-CE val-unseen episodes, while allowing it to navigate without a large language model planner at test time. The method, H2E, turns successful agent sessions into training examples for a model that converts visual and language inputs into actions. It excludes training penalties for steps moving away from the goal and labels visited states within 1 m of the goal as stopping points. The findings are presented in a preprint and concern the specified navigation benchmark.