LongNav-R1: Horizon-Adaptive Multi-Turn RL for Long-Horizon VLA Navigation

Researchers developed LongNav-R1, a multi-turn reinforcement learning framework that teaches Visual-Language-Action models to navigate through continuous conversation with their environment. Unlike single-turn methods, it allows robots to reason about past interactions and future outcomes while learning directly from online exploration. The system improved the Qwen3-VL-2B model's navigation success rate from 64.3% to 73.0% using just 4,000 rollout trajectories, demonstrating superior sample efficiency over state-of-the-art methods. All code will be open-sourced.