VLA Gallery / All models / Navigation

InternVLA-N1

Shanghai AI Laboratory (Intern Robotics) · 2025-07

VLANavigation8.39Bdiffusion

An asynchronous dual-system navigation foundation model where a VLM-based System 2 outputs in-image pixel goals·latent plans at low frequency and a diffusion-based System 1 generates high-frequency trajectories conditioned on them.

Paper

InternVLA-N1: An Open Dual-System Vision-Language Navigation Foundation Model with Learned Latent PlansReport

Overview

Category

Navigation

Type

VLA

Released

2025-07

Organization

Shanghai AI Laboratory (Intern Robotics)

Open weights

Yes

Parameters

8.39B (HF Preview safetensors) / 8.38B (HF DualVLN); System 2 is Qwen2.5-VL 7B

Architecture

Backbone LLM

Qwen2.5-VL-7B (System 2)

Vision encoder

Qwen2.5-VL vision encoder (S2) + DepthAnything RGB encoder (S1)

Action head

S2 pixel goal·latent plan → S1 diffusion trajectory (DualVLN 32 waypoints) (diffusion)

Chunk

Trajectories (32 dense DualVLN waypoints)

Control rate

S2 2 Hz · S1 30 Hz · control 200 Hz

Architecture (original summary)

System 2: Qwen2.5-VL-7B takes past·current RGB and instructions and outputs an in-image pixel goal (or learned latent plan tokens)·stop·turn in place at 2 Hz / System 1: NavDP-style diffusion policy (DepthAnything RGB encoder, with critic; DualVLN uses DiT 12 layers·hidden 384) generates short-horizon trajectories at 30 Hz (DualVLN 32 dense waypoints) → MPC tracking, low-level controller 200 Hz

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Report: S2 SFT 1 epoch; DualVLN paper: S2 14,000 steps·batch 128 (lr 2e-5), S1 15,000 steps·batch 128 (lr 1e-4)

GPU/TPU-hours

Not disclosed

Fine-tuning

Compute for 2-stage joint tuning (S2 frozen, latent plan tokens + S1 asynchronous fine-tuning) not disclosed; with world model extension (Wan2.1 1.3B), optimal R2R-CE epochs 4→2

Data type

Simulation·synthetic

Data scale

InternData-N1: 3,154 scenes·4,839.9 km·1,344 hours·53.5M images·0.8M instructions (VLN-N1 3,000+ scenes·40M images·600k instructions; VLN-CE 332,179 episodes·856 scenes; VLN-PE 8,679 episodes); 5M+ S2 alignment samples

Main datasets

InternData-N1 (VLN-N1: Replica·MP3D·Gibson·3D-Front·HSSD·HM3D; VLN-CE: R2R·RxR·EnvDrop·ScaleVLN replay; VLN-PE: InternUtopia physics sim)

Tasks & evaluation

Technical report: R2R-CE val-unseen SR 55.4% (S2)·58.2% (S1+S2), RxR-CE SR 49.5%·53.5%; zero-shot on VLN-PE (H1) improves SR by 17%+ over models trained on VLN-PE; DualVLN: R2R-CE SR 64.3%·SPL 58.5%, RxR-CE SR 61.4%·SPL 51.8%, VLN-PE R2R unseen SR 51.6%; real-world office·restaurant·street, etc.

Robot

Turtlebot4 wheeled, Unitree Go2 quadruped, Unitree G1 humanoid; RealSense D455 (tilted 15° down)

Inference

Inference device & speed

1 remote RTX 4090, GPU memory about 20GB; S2 trajectory-token KV cache reuse cuts about 1.1 s→0.7 s; S1 with TensorRT does 32 trajectories in about 0.03 s; S2 2 Hz·S1 30 Hz

Inference memory

About 20GB

Weights bf16 (calc.)

16.8 GB

Where it runs

Remote server (RTX 4090)

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (SFT·diffusion) only; S1 critic is supervised

Base control

Wheeled·Quadruped·Humanoid legged (Waypoints → tracking controller) · S2 pixel goal·latent plan → S1 trajectory output → converted to world coordinates via odometry and tracked with MPC

Notes

Key features

An asynchronous dual-system navigation foundation model where a VLM-based System 2 outputs in-image pixel goals·latent plans at low frequency and a diffusion-based System 1 generates high-frequency trajectories conditioned on them. Trained on sim-only InternData-N1 (53.5M images·4,839.9 km), it achieves state of the art on VLN-CE·VLN-PE and transfers zero-shot to wheeled·quadruped·humanoid real robots.

Notes

Announcement based on creation of the HF InternVLA-N1-Preview repository (2025-07-21); technical report 2025-09-11, DualVLN paper arXiv 2512.08186 (2025-12); HF variants: Preview (synchronous, about 2 Hz) → wo-dagger (renamed first official release) → w-NavDP*·DualVLN (recommended, RGB only); training GPU·time Not disclosed; data scale stated variously as 50M+/53M/53.5M; license CC BY-NC-SA 4.0; training code released 2025-12. Based on: NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance (arXiv 2505.08712) — the report states System 1 inherits a structure similar to NavDP

Verification

Confirmed