ABot-N1 reports 95.4% indoor navigation via slow-fast split

A new arXiv paper posted this month claims a fairly large jump on the harder end of indoor and urban robot navigation, and the architectural choice behind the number is more interesting than the number itself. The system, described in an arXiv preprint titled "ABot-N1: Toward a General Visual Language Navigation Foundation Model", reports a 95.4% success rate on complex indoor scenes, 92.9% outdoors, and a 35.0% improvement in point-of-interest arrival to 77.3%.

The design worth reading is the decoupling. Rather than one monolithic policy mapping observations to actions, the authors split the model in two. A slow vision-language reasoner does explicit Chain-of-Thought over the scene and produces what they call a pixel goal, and a fast action expert takes that pixel goal plus text guidance to generate continuous waypoints at the native control frequency. The paper argues this compact image-space anchor is a universal interface across five tasks: point-goal, object-goal, poi-goal, instruction-following, and person-following. The pitch is that you get the interpretability and generalization of a large multimodal reasoner without paying its latency at every control step.

Why this matters if you do not build robots: the slow-fast pattern has been drifting through the vision-language-action literature for over a year, and each credible result narrows the gap between what a large model can plan and what a robot can actually do in the loop. If these numbers survive third-party reproduction, teams shipping delivery, inspection, or in-building service robots gain a reference blueprint that spans multiple task types without a task-specific head for each. The authors also open-source new Point-Goal and POI-Goal benchmarks aimed specifically at urban-scale navigation, which is where these systems tend to fall apart.

Take the specifics as reported, not settled. The abstract does not disclose compute cost, model backbones, or the exact training mix, and the newly released benchmarks are new enough that no third party has run against them yet. There is also the usual home-court concern when the group topping a benchmark is the same group releasing it. The "ABot" line traces to work by AMAP CV Lab, described by Yahoo Finance as "Alibaba Group's mapping unit", which shipped the predecessor ABot-N0 earlier this year; the current paper does not tie itself to a specific deployment.

The direction to watch is whether decoupled slow-fast stacks become the default shape of navigation foundation models, or whether end-to-end training reasserts itself once someone throws a bigger training budget at the problem. Right now the decoupled camp is the one publishing the wins.