Urban-scale and embodied navigation still fail in the wild because end-to-end policies conflate long-horizon reasoning with high-frequency control, causing coordinate drift, brittle long-tail semantics handling, and poor interpretability. The core insight of ABot-N1 is that splitting cognition and control — using explicit V-L reasoning to emit compact, pixel-grounded goals, then feeding those anchors to a high-frequency action expert — yields a navigation foundation that is simultaneously more general, robust, and interpretable.
Key Findings
- Decoupled slow–fast architecture: a slow vision-language reasoner performs explicit Chain-of-Thought and produces pixel-space anchor goals; a fast action expert uses textual cues plus those pixel anchors to output continuous waypoints. This separation reduces drift and preserves high-rate control.
- Broad task coverage with a single interface: the pixel-anchor interface supports point-goal, object-goal, POI-goal, instruction-following, and person-following without task-specific policy changes—so the same model family generalizes across embodied navigation tasks.
- Strong empirical gains at scale: ABot-N1 sets new state-of-the-art marks in urban-scale navigation, increasing POI arrival by 35.0% (to 77.3%) and achieving 95.4% / 92.9% success rates in complex indoor and outdoor scenes, demonstrating robustness across diverse benchmarks.
- Interpretability and diagnostics: pixel-grounded anchors paired with linguistic traces provide human-readable reasoning steps and compact failure modes for easier debugging compared with black-box action policies.
Who it's for and tradeoffs
Great fit if you need a single navigation backbone that generalizes across indoor/outdoor and multiple goal modalities, and you value interpretability and robustness in long-horizon urban scenes. Look elsewhere if your deployment strictly requires minimal latency on devices that cannot afford a slower reasoning loop, or if you need a purely end-to-end learned policy without explicit intermediate representations.
Where it fits
ABot-N1 positions itself between monolithic end-to-end navigation policies (which map observations directly to actions) and heavyweight modular stacks (separate SLAM + planning + perception). It provides an interpretable, compact interface (pixel anchors) that keeps planning lightweight while enabling V-L reasoning to handle semantic complexity and long-range instructions.
How it works (high level)
The slow V-L reasoner ingests visual observations and textual cues, executes explicit multi-step reasoning (Chain-of-Thought) to identify image-space anchor points as subgoals, and emits linguistic traces describing intent. The fast action expert runs at control frequency, fusing pixel anchors and language hints to generate continuous waypoints and control commands. New Point-Goal/POI-Goal benchmarks are released open source to evaluate urban-scale performance and accelerate comparison.