MemVLN

Checkpoint for MemVLN: Streaming Episodic Memory for Vision-Language Navigation, using FRME memory consolidation.

This repository contains memvln.safetensors, the supplied navigation checkpoint, without conversion or re-export. Use it with the MemVLN source code and the matching Qwen3-VL-4B-Instruct and StreamVGGT base weights.

Checkpoint contents

The 2,628,823,872-byte file contains 571 tensors: 504 LoRA tensors, 56 memory tensors, 10 geometry-merger tensors, and the language-model embedding matrix. It contains 514 FP32 tensors and 57 BF16 tensors. The file is a partial checkpoint in MemVLN's parameter naming scheme; it is not a standalone Transformers or PEFT from_pretrained repository.

• File integrity: SHA256SUMS.
• Tensor names, shapes, dtypes, and component counts: checkpoint_inventory.json.
• Portable evaluation configuration: release_configs/eval_release.yaml.
• Configuration/source provenance: release_info.json.
• Data: MemVLN-data.
• Source: MemVLN.
• Paper: arXiv:2610.35066.

Download

hf download Yannnnnnnn/MemVLN memvln.safetensors SHA256SUMS \
  --revision 9d6c106ba1737aa4d06bfb939a0f4c20a15ccecd \
  --local-dir ./checkpoints/MemVLN
(cd ./checkpoints/MemVLN && shasum -a 256 -c SHA256SUMS)

The checkpoint can be downloaded without logging in. The immutable checkpoint revision is recorded in artifacts.json.

Base weights

Download these immutable official releases into the MemVLN source root:

hf download Qwen/Qwen3-VL-4B-Instruct \
  --revision ebb281ec70b05090aa6165b016eac8ec08e71b17 \
  --local-dir ./pretrained/Qwen3-VL-4B-Instruct
hf download lch01/StreamVGGT checkpoints.pth \
  --revision f9ba55b1955bc4f34337c51142158a9aa2862c7f \
  --local-dir ./pretrained/StreamVGGT

See BASE_MODELS.md and artifacts.json for published hashes and provenance. The official StreamVGGT file is named checkpoints.pth; the reference evaluation used the local name StreamVGGT-1B.pth. The published file is a plain PyTorch state dictionary.

Evaluation

Use source revision 4ea586f40a3d18c67454a3e9b7d86b9d72b288eb (tag v0.1.0) with configs/eval_release.yaml. The included release_configs/ directory captures the same configuration with portable model/output paths. It is not a Transformers config.json.

The recipe selects seed 42, eight shards, BF16 memory parameters, FlashAttention 2, nearest 16 observations, FRME with 100 frames and 100 tokens per frame, 300 retrieved memory tokens, grounded-memory caching, fast_habitat=false, and no video output.

After extracting the dataset archive, run from the MemVLN source root:

export MEMVLN_SCENES_DIR=/path/to/scene_datasets
export MEMVLN_R2R_DATA_ROOT="$PWD/datasets/vln/r2r"
export MEMVLN_RXR_DATA_ROOT="$PWD/datasets/vln/rxr"
export MEMVLN_RXR_GT_ROOT="$PWD/datasets/vln/rxr"
EVAL_NOTE="release_frme_$(date +%Y%m%d_%H%M%S)_$$"

bash split_eval/launch_eval_split.sh \
  --gpus 8 \
  --config eval_release \
  --model "$PWD/checkpoints/MemVLN/memvln.safetensors" \
  --epoch 0 \
  base_dir="$PWD/outputs" \
  note="$EVAL_NOTE" \
  llm.cfg_path="$PWD/pretrained/Qwen3-VL-4B-Instruct" \
  vision3d.ckpt_path="$PWD/pretrained/StreamVGGT/checkpoints.pth"

python split_eval/merge_results.py \
  --dir "$PWD/outputs/Qwen3-VLNCE_${EVAL_NOTE}/eval_results/R2R/epoch_0" \
  --expected-splits 8 --expected-episodes 1839
python split_eval/merge_results.py \
  --dir "$PWD/outputs/Qwen3-VLNCE_${EVAL_NOTE}/eval_results/RxR/epoch_0" \
  --expected-splits 8 --expected-episodes 3669

Matterport3D scene meshes are obtained separately under the dataset's access terms. The archive provides the evaluation episode and RxR GT files.

Reported paper results

Dataset	Split	NE (m)	SR (%)	SPL (%)	nDTW (%)
R2R VLN-CE	Val-Unseen	5.19	55.0	50.5	—
RxR VLN-CE	Val-Unseen, English guide	5.59	55.4	48.4	64.6

Release

The checkpoint accompanies the MemVLN v0.1.0 source release. Base-model revisions, file hashes, and the evaluation configuration are recorded in the linked metadata.

Terms and citation

See LICENSE.md and THIRD_PARTY_NOTICES.md. The root source-code MIT license does not by itself license these weights or their training data for unrestricted use.

@misc{wang2026memvln,
  title  = {MemVLN: Streaming Episodic Memory for Vision-Language Navigation},
  author = {Wang, Yan and Huang, Jiangyong and Huang, Siyuan and Jia, Baoxiong},
  year   = {2026},
  eprint = {2610.35066},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url    = {https://arxiv.org/abs/2610.35066}
}