SimpleVLN: Native MLLMs Achieve Navigation SOTA Without Extra Modules

Problem: Can Native MLLMs Navigate Without Specialized Modules?

Vision-language navigation (VLN) requires robots to follow natural language instructions through continuous environments. Most existing methods add spatial reasoning modules, graph structures, or hierarchical decision systems. The authors ask whether a native multimodal large language model (MLLM) can achieve strong navigation solely through clever decision design and training, without modifying the model architecture or adding navigation-specific components.

Method: SimpleVLN Framework

Observation Selection: Global + Local Sampling

Each model call is treated as an independent decision. Inputs include the navigation instruction, current observation, and a filtered history. Instead of accumulating all past observations (which lengthens input) or keeping only recent frames (which loses early route information), SimpleVLN uses a combined strategy: uniformly sample from early global observations and always retain the latest local frame. Ablation shows feeding all history does not improve performance; uniform global sampling works best, and adding more local frames yields no further gains.

Action Chunking

The model predicts an ordered sequence of basic actions (an action chunk) per decision, executed sequentially by the environment. On Qwen3-VL 4B, action chunking raises success rate from 45.6% to 53.6%, demonstrating that how information is organized for a decision and how actions are output both affect continuous navigation.

Three-Stage Training Pipeline

Stage 1: Supervised Fine-Tuning (SFT) on Expert Demonstrations

Model learns from expert trajectories on R2R, RxR, and ScaleVLN datasets.

Stage 2: DAgger with Deviation-Based Filtering

To handle distribution shift when the robot deviates from expert paths, the team employs DAgger. They compare policy-generated trajectories with ground truth: large deviations use expert trajectories for supervision; small deviations retain the model's own trajectory. These are converted to decision-level supervision samples and merged with original expert data. After DAgger on combined R2R, RxR, ScaleVLN data, success rates on R2R-CE rise from 57.9% to 65.4%, and on RxR-CE from 55.3% to 63.4%.

Stage 3: Reinforcement Learning (RL) with Task-Level Rewards

SFT and DAgger rely on per-step action matching, which cannot reward a valid but different path that reaches the goal. The authors introduce RL (DAPO) on newly generated navigation tasks with diverse start/goal pairs and language instructions. The model interacts via action chunks until task completion; reward is 1 for success (stopping near target), 0 otherwise. This allows the model to learn from its own exploration without mimicking expert routes. On Qwen3-VL 4B, RL boosts R2R-CE from 65.4% to 73.8% (+8.4 pp) and RxR-CE from 63.4% to 75.1% (+11.7 pp).

Results Across Models and Benchmarks

Qwen3.5 9B achieves 76.7% (R2R-CE) and 80.1% (RxR-CE) on unseen environments, outperforming Qwen-RobotNav 8B by 4.6 and 3.6 percentage points respectively.

Inference latency: SimpleVLN on Qwen3-VL 4B runs at 293 ms per action chunk on NVIDIA A100, comparable to streaming method StreamVLN.

Real-World Deployment

The framework is deployed on a quadruped robot dog. It executes multi-step commands such as "walk down the corridor, turn left, enter the first room" and "walk forward, turn left, continue forward, stop beside the yellow pillar", demonstrating strong generalization in physical environments.

Key Takeaways

Native MLLMs possess strong latent navigation capabilities that can be unlocked via observation selection (global+local sampling) and action chunking.

A three-stage training pipeline (SFT → DAgger with deviation filtering → RL with task-level rewards) progressively closes the sim-to-real gap and improves error recovery.

No architectural changes to the MLLM are required; the approach is model-agnostic and efficient.

Paper: SimpleVLN: Towards Simple and Frontier Vision-Language Navigation With Native MLLMs (arXiv:2610.23203v1) Project page: https://simplevln.github.io/ Code: https://github.com/simplevln/SimpleVLN