ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied Navigation

TL;DR

ABot-N0 proposes a unified Vision-Language-Action foundation model for versatile embodied navigation by coupling a Cognitive Brain (LLM) with a Flow-Matching Action Expert in a Brain-Action hierarchy. A three-layer Data Engine (7,802 high-fidelity scenes; 16.9M trajectories; 5.0M reasoning samples) fuels cross-task generalization across Point-Goal, Object-Goal, Instruction-Following, POI-Goal, and Person-Following, achieving state-of-the-art results on seven benchmarks. The Agentic Navigation System augments ABot-N0 with a Planner, hierarchical Topo-Memory, and a Closed-loop Self-Reflector for long-horizon, robust behavior in dynamic real-world settings, including deployment on a Unitree Go2 with onboard inference. Together, these components demonstrate strong cross-task generalization, scalable data synthesis, and practical real-world robustness, enabling socially aware, long-horizon autonomous navigation across indoor and outdoor environments.

Abstract

Embodied navigation has long been fragmented by task-specific architectures. We introduce ABot-N0, a unified Vision-Language-Action (VLA) foundation model that achieves a ``Grand Unification'' across 5 core tasks: Point-Goal, Object-Goal, Instruction-Following, POI-Goal, and Person-Following. ABot-N0 utilizes a hierarchical ``Brain-Action'' architecture, pairing an LLM-based Cognitive Brain for semantic reasoning with a Flow Matching-based Action Expert for precise, continuous trajectory generation. To support large-scale learning, we developed the ABot-N0 Data Engine, curating 16.9M expert trajectories and 5.0M reasoning samples across 7,802 high-fidelity 3D scenes (10.7 $\text{km}^2$). ABot-N0 achieves new SOTA performance across 7 benchmarks, significantly outperforming specialized models. Furthermore, our Agentic Navigation System integrates a planner with hierarchical topological memory, enabling robust, long-horizon missions in dynamic real-world environments.

Figures (21)

• Figure 1: ABot-N0: A unified VLA foundation model for versatile embodied navigation. Powered by a massive dataset of 16.9M expert trajectories and 5M reasoning samples across diverse environments, the model achieves a "Grand Unification" across five core navigation tasks. It establishes new state-of-the-art performance across 7 challenging benchmarks and is successfully deployed in complex, dynamic real-world agentic navigation systems.
• Figure 2: The Architecture of ABot-N0. The model adopts a hierarchical "Brain-Action" design. The Universal Multi-Modal Encoder unifies heterogeneous inputs (RGB observations, visual history, and goal specifications) into a shared token sequence. The Cognitive Brain (i.e., LLM) encodes these tokens and supports dual-mode operation: a Reasoning Head for high-level semantic understanding and an Action Head for motion planning. The Action Expert employs Flow Matching to generate trajectory distributions, enabling generalization across five navigation tasks.
• Figure 3: High-Fidelity 3D Scene Ecosystem. Our data engine integrates diverse indoor and outdoor environments. Left: Indoor scenes span residential spaces to large-scale public venues (offices, malls, transit stations). Right: Outdoor scenes include real-world scans of intersections and parks, alongside the dynamic virtual city SocCity. All scenes are annotated with traversable navigation graphs for collision-free trajectories generation.
• Figure 4: 3D Scene Ecosystem Statistics. Our collection comprises 7,802 high-fidelity 3D scenes, covering 6.25 $\bm{km^2}$ of indoor environments (offices, malls, stations, homes) and 4.42 $\bm{km^2}$ of outdoor environments (intersections, parks, city). Navigation graphs totaling 384,754 meters enable collision-free trajectory synthesis across diverse spatial scales—from compact residential units to expansive transit hubs and urban environments.
• Figure 5: Point-Goal Navigation Data Pipeline. Our 4.0M point-goal trajectory dataset aggregates three complementary streams: (Top) 2.0M pseudo-trajectories from first-person internet videos via 3D structure recovery and metric alignment; (Middle) 1.7M synthetic trajectories from high-fidelity 3D scenes with annotated navigation graphs; (Bottom) 340K real-world robot demonstrations from heterogeneous datasets, providing ground-truth physical dynamics.