SimpleNav Open-Sourced: One Framework Covers the Full Navigation VLA Pipeline Across 6 Benchmarks and Three Embodied Scenarios
Navigation vision-language-action (VLA) models take language instructions and continuous visual observations as input, generate executable action sequences, and drive embodied platforms such as drones, unmanned ground vehicles, and mobile robots to complete navigation tasks. Conducting navigation VLA research requires not only building efficient models but also a complete pipeline that connects data processing, training optimization, and closed-loop evaluation. Reducing repeated adaptation work across different data, models, and evaluation benchmarks has become a key issue for improving navigation VLA R&D efficiency.
In practice, datasets differ markedly in coordinate systems, camera configurations, trajectory distributions, and action representations, and different tasks use different simulation environments, interaction interfaces, and evaluation protocols. Whenever researchers switch datasets, model architectures, or evaluation benchmarks, they often have to redo data conversion, interface adaptation, training configuration, and evaluation alignment. These repetitive, scattered engineering steps not only raise experiment costs but also hold back the efficiency of model development, method reproduction, and performance comparison.
To address this, the THUNLP lab at Tsinghua University, AI9stars, OpenBMB, and HITDIP jointly developed SimpleNav, an open-source full-pipeline framework for navigation VLA research. It connects key stages including heterogeneous data processing, modular model construction, unified training, and closed-loop evaluation, providing standardized onboarding and validation flows for different datasets, models, and tasks. The first open-source release supports 6 evaluation benchmarks — OpenFly, TravelUAV, AerialVLN, R2R-CE, RxR-CE, and EVT-Bench — covering three embodied navigation scenarios: aerial navigation, indoor navigation, and moving-target tracking, helping researchers reproduce, develop, train, and evaluate navigation VLA models more efficiently.
Figure 1: Examples of SimpleNav closed-loop evaluation on aerial navigation and indoor navigation tasks
➤ Open-source link
🔗 https://github.com/OpenBMB/SimpleNav
The navigation VLA challenge: why does every new scenario require re-adapting the experiment pipeline?
The problem is not only that the stages differ, but that data, models, training, and evaluation are often tightly coupled. A dataset's coordinate system and action representation determine the model's input and output form; the model structure affects training objectives and optimization; and the actions the model generates must match the simulation environment's control interface and evaluation protocol. As a result, switching datasets, tasks, or embodied platforms often requires coordinated changes across multiple stages, making it hard to reuse an existing experiment pipeline.
As navigation foundation models, datasets, and embodied platforms proliferate, navigation VLA research is moving from a single dataset, scenario, and platform toward unified modeling and generalization validation across data, scenarios, and platforms. In this trend, whether the experiment pipeline can be reused directly affects the efficiency of validating new models and methods, and determines whether results from different experiments can be compared reliably under unified conditions.
To this end, SimpleNav turns common engineering capabilities — data conversion, model integration, training configuration, and closed-loop evaluation — into standardized, reusable components and workflows, reducing repeated development when datasets, models, or benchmarks change, so researchers can devote more effort to core research questions such as model architecture, training strategies, multi-dataset mixed training, and cross-scenario generalization.
What "Simple" means: making projects easier to reproduce, develop, migrate, and sustain
The "Simple" in SimpleNav does not mean simplifying research questions or sacrificing framework capability. Rather, it lowers the engineering barrier to navigation VLA research through unified data specifications and modular engineering. Around how an experiment can be reliably reproduced, how a method can be developed quickly, how a capability can be migrated to new tasks, and how existing results can support follow-up research, SimpleNav makes "simple" concrete in four aspects: reproduction, development, migration, and continuity.
The first SimpleNav open-source release: connecting data, models, training, and evaluation for navigation VLA
The first SimpleNav open-source release is organized around four core stages: the data side converts, augments, and validates navigation data from different sources into data ready for training; the model side provides a runnable reference implementation and organizes key components into replaceable modules; the training side uniformly supports single-dataset, joint multi-dataset, and distributed training; and the evaluation side connects the closed-loop environments and metric systems of different benchmarks, completing model execution, trajectory logging, and experiment statistics in a unified flow. The data, model configurations, checkpoints, and evaluation trajectories involved can all be reused for subsequent experiment reproduction, analysis, and further research.
Figure 2: The SimpleNav overall framework
Try SimpleNav: run the full navigation VLA experiment pipeline in four steps
01 Data processing: bringing heterogeneous navigation data into a unified training pipeline
Navigation datasets differ noticeably in image organization, trajectory structure, state recording, temporal sampling, and action representation. SimpleNav consolidates these heterogeneous data into a unified LeRobot v3 data format, so that images, states, actions, metadata, and other information can all be read through the same training interface.
During data conversion, SimpleNav also handles task differences in coordinate systems, timestamps, camera order, trajectory segmentation, normalization statistics, and action representation, and uses data validation to keep bad data out of training. Each benchmark's original coordinate and action definitions are preserved by the corresponding interface, without changing the task's own execution and evaluation conventions.
SimpleNav also provides preprocessing tools for trajectory augmentation, image capture and interpolation, and image cache generation, forming a continuous data pipeline from raw data cleanup and augmentation to training use.
Figure 3: SimpleNav data preprocessing and Openfly trajectory augmentation examples
02 Model construction: a reference implementation that runs out of the box, with key modules that can be swapped independently
SimpleNav provides a reference model that has already completed training and closed-loop validation, covering the full path from visual and language input to continuous navigation action output. On this basis, the vision-language backbone, history information processing, and action prediction modules are organized as independent components, so researchers can replace only the part they want to study while reusing the existing data organization, training pipeline, and benchmark evaluation.
This componentized approach supports direct experiment reproduction and makes it easy to run ablations and comparisons around new vision-language backbones, long-history modeling, memory mechanisms, or action heads, concentrating engineering changes on the variables that actually need study.
03 Unified training: single-dataset, joint multi-dataset, and distributed training
For a single benchmark, SimpleNav can directly use the corresponding data for standalone training and experiment reproduction; further, the same training framework also supports mixed training over multiple navigation datasets, providing a foundation for cross-data, cross-scenario, and unified navigation model research.
At the compute level, the training entry point covers both single-machine multi-GPU and multi-machine distributed runs. Whether changing model structure, data combinations, or training scale, experiments can follow the same training organization and simultaneously save the effective configuration, data statistics, logs, and checkpoints, making continued training, ablation, and comparison easy.
04 Closed-loop evaluation: unified environment interaction, trajectory logging, and metric computation
Navigation models ultimately need to be tested through continuous environment interaction. SimpleNav connects the closed-loop execution of different benchmarks into a unified evaluation framework: the model reads the current observation and outputs an action, the environment advances according to the corresponding task protocol, and the new visuals and states are returned to the model until the full episode ends. Each task's own action definitions, stopping conditions, and metric computation remain the responsibility of the corresponding evaluation interface.
The evaluation process simultaneously records full trajectories, per-episode results, and aggregate metrics. Beyond final numbers such as success rate, researchers can return to specific trajectories to analyze phenomena such as veering off course, stopping too early, or local hesitation, providing evidence for further improvements to models, data, and training strategies.
Figure 4: Examples of SimpleNav closed-loop evaluation on selected benchmarks
A unified framework with no special optimizations: how does it perform? 6 benchmarks validate cross-scenario performance
To verify that SimpleNav can not only connect cross-scenario experiment pipelines but also provide a competitive model baseline for navigation VLA research, we used Qwen3.5-VL as the unified vision-language backbone and completed model training and closed-loop evaluation on 6 benchmarks. Apart from the necessary data and task interface adaptation, no benchmark-specific performance optimization was applied. Therefore, these results reflect the baseline performance of the SimpleNav reference model under a unified experiment framework, not the performance ceiling after per-task deep tuning.
These results show that SimpleNav can connect aerial navigation, indoor navigation, and moving-target tracking tasks under one framework, and that its reference model achieves stable, competitive performance without benchmark-specific optimization. This provides a reproducible, extensible unified framework with a strong performance starting point for subsequent research on model architecture improvements, multi-dataset training, and cross-scenario generalization.
From a unified pipeline to sustained cross-data, cross-platform research
The SimpleNav project open-sources the complete navigation VLA pipeline from data processing and model construction to training and closed-loop evaluation: heterogeneous navigation data can be onboarded uniformly, key model components can be swapped flexibly, single-dataset, multi-dataset, and distributed training are supported, and closed-loop evaluation covers multiple aerial and indoor navigation benchmarks. Accompanying models, checkpoints, experiment results, and reproduction documentation are also released, so that future navigation VLA research can go from data preparation to final validation within the same framework.
Going forward, SimpleNav will continue to expand to more datasets, more models, joint multi-dataset training, more embodied platforms, and real-world environments, moving navigation VLA research toward a more open, reusable, and sustainable research system.
Make navigation VLA simple