RSI-Nav: recursive self-improvement for navigation agents hits 68.6% SR on R2R-CE
OpenGVLab · hf · 2026-10-14
OpenGVLab and collaborators introduce RSI-Nav, a data-centric recursive self-improvement framework for vision-language navigation:
- Motivation: human annotation is costly, and existing synthesis pipelines annotate full corpora with a fixed model, letting supervision lag behind the evolving policy.
- Method: unify instruction following (VL-to-action) and instruction generation (V-action-to-language) in one shared model; after SFT, process unlabeled trajectories in shards — generate instructions, execute them, evaluate goal completion and path efficiency, keep qualified samples, and update immediately for policy-synchronized supervision.
- Spatial understanding data is mixed in to boost spatial reasoning.
RSI-Nav reaches 68.6% SR on R2R-CE and 72.4% SR on RxR-CE, with improved instruction generation and modest spatial gains.
More from Embodied
- Nature: smart glasses are heading for a major privacy crisis — marigo · 2026-10-14
- Impatient for JetKVM Mini, dev buys LuckFox KVM to let agents control his Mac Mini — TheZachMueller · 2026-10-14
- China's electric autonomous tractor T70 goes viral, already deployed in Hebei — lukas_m_ziegler · 2026-10-14
- Nomagic teases more updates as warehouse robots tackle unexpected scenarios — m_wulfmeier · 2026-10-14
- NTU team has Claude write robot control code, hitting 70.24% success in sim — aigclink · 2026-10-14
- Rabbit R1 called 'ahead of its time' as modders plan to hack the once-mocked AI device — SimonBalmain · 2026-10-14