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

AMAP CV Lab arXiv

TL;DR

Motivations & Innovations

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.

• Point-Goal. The agent must reach precise metric coordinates defined in a local frame, serving as the foundational primitive for robust locomotion and obstacle avoidance.
• Object-Goal. The agent actively searches for and navigates to a specific object category in unseen environments, requiring sophisticated semantic reasoning and multimodal integration.
• Instruction-Following. The agent must execute long-horizon, complex natural language paths, focusing on the rigorous alignment between linguistic input and sequential action execution.
• POI-Goal. requires the agent to identify specific Points of Interest (POIs) and precisely navigate to their physical entrances, bridging outdoor and indoor environments while addressing the last-meters navigation challenge.
• Person-Following. This involves real-time tracking of dynamic human targets, representing a critical social capability for human-robot interaction.

Model Architecture

Encoder

LLM Backbone

a pre-trained LLM (Qwen3-4B)

Action Expert

Data Recipe

High-Fidelity 3D Scene Ecosystem

-> generalize across diverse physical environments.

high-fidelity 3D scenes.

closed-loop simulation.

Indoor Environments: From Domestic to Public

Residential Scenes (Home)

• HM3D
• InteriorGS

Public Spaces (Office, Mall, Station)

Outdoor Environments: Static Scans and Dynamic City

Real-World Scans (Intersection, Park)

• Intersections
• Parks

Dynamic Virtual City (SocCity)

ABot-N0 Trajectories Dataset

Point-Goal Trajectories (4M)

Deriving Pseudo-Trajectories from Internet Videos (2M). We collectfirst-person urban exploration videos to capture the long-tail distribution of global architectural styles and weather conditions.

• Structure Recovery: utilize $\pi^3$ for dense 3D reconstruction from monocular streams.
• Metric Alignment: apply MoGe to resolve scale ambiguity and align the reconstructed paths with real-world metric units.
• Episode Synthesis: sample diverse point-goal pairs along the camera’s motion manifold to generate kinematically consistent navigation samples

Synthesis from High-Fidelity 3D Scenes (1.7M). 3D scenes and dynamic SocCity.

Point-goal episodes are generated by sampling reachable coordinate pairs (s, g) and computing optimal paths using path-finding algorithms. Crucially, we incorporate recovery trajectories by initializing agents in off-path or near-collision states, forcing the model to learn error-correction behaviors essential for real-world robustness.

https://arxiv.org/abs/2511.21135

https://amap-eai.github.io/SocialNav/

Integration of Heterogeneous Robot Demonstrations (340K). diverse real-world robotic platforms (e.g., SCAND [22], HuRoN [18], Recon [41], and CityWalker [36] teleoperation data).

Instruction-Following Trajectories

Object-Goal Trajectories

HM3D and OVON

OVON-sub

InteriorGS

ABot-N0 Reasoning Dataset

General Navigation Skills

Navigation Chain-of-Thought (CoT):

Instruction-Following Reasoning

Object-Goal Reasoning

POI Grounding

General VQA (1.1M)

• VQA
• Embodied Question Answering (EQA)

Training Recipe

Deploy

Real-world Deployment

• Unitree Go2 quadrupedal robot with 12 DoF

  • Sensors:
    • RTK-GNSS for global localization
    • 3 RGB cameras (forward/left/right, H120°, V90°)
    • Unitree 4D LiDAR L2

• NVIDIA Jetson Orin NX (157 TOPS, 16GB RAM), in which perception and execution tasks are performed, including the real-time inference of the optimized VLA model.

• 2Hz VLA inference and 10Hz closed-loop control