ABot-N0

AMAP CV Lab (Alibaba) · 2026-02

VLANavigation4Bflow matching

Unifies 5 navigation tasks (point·object·instruction·POI·person following) in one VLA with a Brain-Action architecture combining an LLM Brain and a flow matching Action Expert, trained with a data engine of 16.9M expert trajectories·5.0M reasoning samples and SAFE-GRPO RL.

View architecture diagram in the gallery Paper ↗

Paper

ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied NavigationarXiv 2602.11598

Overview

Category: Navigation
Type: VLA
Released: 2026-02
Organization: AMAP CV Lab (Alibaba)
Open weights: No
Parameters: 4B (based on Qwen3-4B + edge version SigLIP-B/16 93M; overall figure not stated)

Architecture

Backbone LLM: Qwen3-4B
Vision encoder: ViT (edge version SigLIP-B/16 93M)
Action head: flow matching action expert: 5 BEV waypoints (x, y, θ) + reasoning head (flow matching)
Chunk: 5 waypoints
Control rate: VLA 2 Hz · controller 10+ Hz
Architecture (original summary): General multimodal encoder (ViT; panorama left·front·right or front only, episodic visual memory, text·point goals) + Cognitive Brain Qwen3-4B (reasoning head·action head) + flow matching Action Expert → 5 local BEV waypoints (x,y,θ); onboard VLA 2 Hz (edge version SigLIP-B/16 93M·token merging), low-level Neural Controller >10 Hz

Training

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed (3 stages: perception warm-up → unified sensorimotor SFT → SAFE-GRPO)
GPU/TPU-hours: Not disclosed
Fine-tuning: Stage 3 SAFE-GRPO (Brain frozen, only the Action Expert trained with RL on SocCity expert trajectories · composite rewards): accelerator and time not disclosed
Data type: Real-robot teleoperation / real-robot autonomous driving logs / web video / simulation·synthetic / web vision-language
Data scale: 16.9M expert trajectories (PointGoal 4.0M: 2.0M pseudo-trajectories from internet video, 1.7M synthetic, 340K real robot / instruction following 2.8M / ObjectGoal 3.6M / POI 2.5M (Wan2.1-I2V generated video) / person following 4.0M) + 5.0M reasoning samples (incl. 1.1M general VQA); 7,802 3D scenes, 10.7 km² (paper)/10.3 km² (README)
Main datasets: VLN-CE R2R·RxR; HM3D·HM3D-OVON; InteriorGS; SocCity (in-house virtual city); SCAND; HuRoN; RECON; CityWalker; internet video (π³·MoGe reconstruction)
Tasks & evaluation: R2R-CE val-unseen SR 66.4%·SPL 63.9, RxR-CE SR 69.3%·SPL 60.0; HM3D-OVON val-unseen SR 54.0%; SocNav closed-loop SR 88.3% (CityWalker 47.8%); CityWalker open-loop MAOE 11.2 (15.2); BridgeNav POI SR@0.3 m 88.68%; EVT-Bench STT SR 86.9%
Robot: Unitree Go2 X quadruped (3 RGB cameras about 270°, 4D LiDAR L2, RTK-GNSS)

Inference

Inference device & speed: On Unitree Go2 onboard Jetson Orin NX (157 TOPS, 16 GB): VLA 2 Hz (token compression, 3% performance drop), Neural Controller >10 Hz; agent planner on cloud RTX 4090; memory usage not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): 8 GB
Where it runs: Onboard Jetson Orin NX (planner on cloud RTX 4090)

Reinforcement learning & base

Reinforcement learning: Used (VLA RL post-training) · Stage 3 SAFE-GRPO applies RL to the flow matching Action Expert (Brain frozen); low-level Neural Controller also RL fine-tuned in Isaac Sim
RL evidence: https://arxiv.org/html/2602.11598v1
Base control: Quadruped (Waypoints → tracking controller) · VLA outputs 5 BEV waypoints at 2 Hz → Neural Controller (CE-Nav-based, >10 Hz) generates body velocity (vx,vy,vyaw) → Go2 locomotion control

Notes

Key features: Unifies 5 navigation tasks (point·object·instruction·POI·person following) in one VLA with a Brain-Action architecture combining an LLM Brain and a flow matching Action Expert, trained with a data engine of 16.9M expert trajectories·5.0M reasoning samples and SAFE-GRPO RL. Claims SOTA on 7 benchmarks (R2R-CE SR 66.4%, SocNav closed-loop SR 88.3%, etc.) and runs onboard at 2 Hz on Go2's Jetson Orin NX.
Notes: README·paper make no mention of weight·code release (README has only paper·project page, marked Apache 2.0). Conflicting scene area: paper 10.7 km² vs README 10.3 km². Some documents say 'ABot-NO'. The low-level Neural Controller is a CE-Nav-based VelFlow expert fine-tuned with Isaac Sim RL. Training compute·total parameters not disclosed.
Verification: Corrections applied