Embodied-Navigator-7B-GRPO

Point, Think, Memorize, and Align for Efficient Embodied Navigation

Project Page | Code | Paper

Embodied-Navigator-7B-GRPO is the released navigation policy checkpoint for Embodied-Navigator. It adapts Qwen2.5-VL-7B to continuous vision-language navigation and aligns the policy with Two-Level Group Relative Policy Optimization (GRPO).

The policy observes four egocentric RGB views, an instruction, and compact trajectory memory. At each decision point it determines whether explicit reasoning is useful, selects a camera view, and predicts a 2D pixel waypoint. The complete system projects that pixel into 3D using depth after the model prediction and delegates motion execution to a low-level navigation controller.

Model Details

Field	Value
Base model	Qwen2.5-VL-7B-Instruct
Model family	Vision-language navigation policy
Precision	BF16
Context configuration	128K tokens
Visual input	Four 90-degree RGB views with 360-degree coverage
Policy output	Selective reasoning, view selection, and 2D pixel waypoint
Navigation memory	Anchor-Trajectory Memory with Space-Time Indicators
Post-training	Supervised fine-tuning followed by Two-Level GRPO
Training data	MultiNav-CoT, 90K navigation trajectories

The checkpoint includes navigation-specific tokens and a learned action head. It is intended to be loaded with the custom Qwen2.5-VL implementation in the project repository rather than treated as a generic image-captioning or chat checkpoint.

Method Summary

Embodied-Navigator organizes the policy around four components:

• Point: predict a view and pixel waypoint, then use deterministic geometry for pixel-to-3D projection.
• Think: trigger Chain-of-Thought reasoning only at decision-relevant nodes.
• Memorize: retain critical visual-reasoning anchors and compress routine motion into Space-Time Indicators.
• Align: combine local action advantages with global trajectory advantages through Two-Level GRPO.

Download

hf download UnderTides/Embodied-Navigator-7B-GRPO \
  --local-dir Embodied-Navigator-7B-GRPO

The repository contains approximately 17 GB of BF16 safetensors split across four shards.

Loading the Checkpoint

Clone the project code and load the checkpoint through its navigation-adapted Qwen2.5-VL classes:

git clone https://github.com/ZJU-OmniAI/Embodied-Navigator.git
cd Embodied-Navigator

from src.model.qwen2_5_vl import (
    Qwen2_5_VLForConditionalGeneration,
    Qwen2_5_VLProcessor,
)

model_id = "UnderTides/Embodied-Navigator-7B-GRPO"

processor = Qwen2_5_VLProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
)
model.eval()

The full navigation workflow builds multi-view prompts, maintains Anchor-Trajectory Memory, parses policy outputs, projects predicted pixels into 3D, and executes waypoints through the environment controller. Use the agent and evaluation code in the project repository for end-to-end evaluation.

Example evaluation entry point:

bash scripts/run_evaluate.sh \
  config/ht_dthink_r2r.yaml \
  ./Embodied-Navigator-7B-GRPO \
  runs/embodied_navigator_eval

This evaluation also requires the corresponding Habitat-Lab environment, benchmark episodes, and licensed Matterport3D assets.

Evaluation Results

Results below are reported on validation-unseen splits. NE is lower-is-better; all other metrics are higher-is-better.

R2R-CE Val-Unseen

NE	OS	SR	SPL
3.85	74.5	66.2	58.8

RxR-CE Val-Unseen

NE	SR	SPL	nDTW
4.32	65.7	56.9	72.4

Additional reported findings:

• Adaptive reasoning reaches 66.2% R2R-CE SR with a 26.3% reasoning ratio.
• Anchor-Trajectory Memory reaches 49.8% SR on the long-horizon subset.
• Zero-shot real-world evaluation reaches 60.0% success over 100 blind trials.

See the project page and repository for complete tables, ablations, qualitative trajectories, and deployment videos.

Intended Use

This checkpoint is intended for research on:

• Continuous vision-language navigation
• Embodied vision-language policies
• Selective reasoning and long-horizon memory
• Pixel-grounded action prediction
• Reinforcement-learning alignment for navigation

It is a component of a complete navigation system. Deployment requires the project's prompt construction, memory management, pixel-to-3D projection, localization, and low-level motion-control modules.

Limitations

• The VLM observes RGB, while the complete system still uses depth for post-prediction geometric projection and odometry for memory encoding.
• Results depend on the full evaluation stack and are not reproduced by loading the checkpoint as a standalone generic Transformers pipeline.
• The policy may stop prematurely or hallucinate success when the target leaves all camera views.
• Performance outside the reported navigation domains, sensor configuration, and instruction distribution has not been established.
• The checkpoint does not include licensed Habitat-Matterport3D assets, the full MultiNav-CoT corpus, or the robot localization and planning stack.

Training Data

The policy is trained on MultiNav-CoT, a 90K-trajectory navigation dataset with Chain-of-Thought annotations generated using Gemini 2.5 Flash. A data subset and the processing pipeline are available in the project repository; the complete training corpus is distributed separately.

License

License terms for the released checkpoint have not yet been specified by the authors. Refer to the project repository for future license updates.

Citation

@inproceedings{feng2026embodiednavigator,
  title     = {Embodied-Navigator: Point, Think, Memorize, and Align
               for Efficient Embodied Navigation},
  author    = {Feng, Hongyan and Chen, Sunlai and Liu, Xuanyu and Pan, Miao and
               Xie, Yangfan and Cui, Yuxiang and Zhou, Zhongxiang and
               Xiong, Rong and Zhang, Wenqi and Yin, Jianwei and
               Zhuang, Yueting and Zhang, Xuhong},
  year      = {2026}
}

Authors

Hongyan Feng, Sunlai Chen, Xuanyu Liu, Miao Pan, Yangfan Xie, Yuxiang Cui, Zhongxiang Zhou, Rong Xiong, Wenqi Zhang, Jianwei Yin, Yueting Zhuang, and Xuhong Zhang.