What it is
An embodied navigation framework that lets a VLM point at image pixels, reasons only at key nodes, and aligns planning with dense RL rewards for efficient VLN. Instead of forcing VLMs into atomic actions or 3D regression, it turns navigation into 2D pixel pointing plus SLAM, triggers CoT only at critical nodes, compresses the rest into spatiotemporal memory, and trains with two-level GRPO using step and trajectory rewards.
Key results
- 66.2% SR on R2R-CE val-unseen, beating DualVLN's 64.3%
- 65.7% SR and 72.4 nDTW on RxR-CE val-unseen, above 61.4% and 70.0
- 16.58s/task on one A800 vs StreamVLN 37.47s and DualVLN 41.46s
Aligning VLM actions to their native 2D priors and reasoning sparsely at key nodes can improve both navigation accuracy and runtime efficiency.