Slow-Fast Navigation Model Improves Urban POI Arrival

Visual-language navigation systems have to connect two different problems: semantic reasoning over goals and reactive control under changing observations. Many recent policies compress that path into a single observation-to-action model, which can work on benchmark tasks but leaves two recurring failure modes: accumulated coordinate drift and weak handling of rare or semantically specific targets. ABot-N1 addresses that split directly. The paper proposes a navigation foundation model in which a slow vision-language reasoner decides where the agent should go, while a faster action expert turns that guidance into continuous waypoints.

Core Contribution

The central contribution is the interface between cognition and control. Instead of asking one policy to infer intent, localize a target, and output motion in one step, ABot-N1 emits two forms of guidance: explicit chain-of-thought text and image-space anchors. The slow system produces a target pixel and an affordance pixel; the fast system consumes those anchors, the language trace, and the current visual observation to drive the agent.

That choice is more than an interpretability feature. Pixel goals give the controller a compact spatial target that is shared across point-goal, object-goal, POI-goal, instruction-following, and person-following tasks. The paper argues that this reduces the burden on the low-level policy: it no longer has to recover the entire semantic plan from raw observations at every control step.

Figure 1 summarizes the overall design and evaluation scope: ABot-N1 is trained on 30M samples across five tasks, then tested on two new ABotN benchmarks and three existing navigation benchmarks.

Technical Approach

ABot-N1 uses a slow-fast dual-system architecture. The slow system is a vision-language reasoner that operates at low frequency over historical frames and the task prompt. It produces an explicit reasoning trace and two visual anchors, which become the communication channel to the controller. The fast system runs at the native control frequency. It uses a lightweight vision-language model, integrates the slow system's textual and pixel-level guidance with live observations, and predicts continuous waypoints.

The action expert is not just a small classifier attached to the reasoner. The architecture shown in Figure 2 includes a learnable action query, a QFormer module that attends to hidden states, and an MLP that decodes those queries into waypoints. This separates deliberation from execution: slow reasoning can be expressive without blocking high-frequency control.

The data pipeline is also part of the method. The paper describes a simulation-based data engine covering indoor and outdoor scenes, with trajectories generated from expert and DAgger rollouts. The resulting corpus spans five pretraining tasks, followed by post-training data stratified into safe, critical, danger, and discarded categories. The authors report both pretraining and GRPO post-training, which matters because the claimed safety and benchmark gains depend on more than the architectural split.

Results and Analysis

The strongest reported numbers are in urban-scale navigation. The paper reports POI arrival rising to 77.3%, described as a 35.0% gain, and success rates of 95.4% and 92.9% in complex indoor and outdoor scenes. It also reports leading performance across five benchmarks: ABotN-PointBench, ABotN-POIBench, VLN-CE R2R/RxR, Short-Horizon OVON, and EVT-Bench.

The interpretation is that ABot-N1 is most convincing where the task needs both semantic grounding and local control. POI navigation is a good stress test for that claim because it requires recognizing a long-tail place target, keeping spatial intent stable, and reacting to visual context. A pixel-anchor interface is a plausible mechanism for the reported gains: it gives the action expert a spatially grounded target rather than relying only on a latent instruction embedding.

The evidence is broad rather than minimal. The model is evaluated across five benchmarks and multiple task types, and the training scale is substantial at 30M samples. The paper also introduces ABotN-PointBench and ABotN-POIBench, which directly target point-goal and POI-goal navigation rather than relying only on legacy instruction-following benchmarks.

The caveat is that the visible results still leave practical questions. The method depends on the slow system producing useful pixels; if those anchors are wrong, the fast controller inherits the error. The training recipe is large and multi-stage, so the compute and data requirements may limit reproduction. The reported benchmark coverage is strong for navigation evaluation, but the paper material provided here does not establish how the system behaves under severe perception noise, changing real-world maps, or tight latency budgets.