DynVLA_Training_Data

# DynVLA Training Data This repository provides the processed data assets and released checkpoints used by [DynVLA](https://github.com/yaoyao-jpg/DynamicsVLA) for Navsim training and evaluation. ## Overview The dataset repository contains: - released Navsim metadata pickle files - raw front-camera image archives - VQ-code archives for Stage 2 training - BEV label archives - released pretrained checkpoints used by DynVLA The files in this repository are intended to be unpacked into the directory layout expected by the DynVLA codebase. ## Repository Layout ```text DynVLA_Training_Data ├── BEV │ └── BEV_maps.tar.gz ├── image │ └── nuplan_images_256x144.tar.gz.part-* ├── image_vq │ └── vq_codes_low_res_corrected_merge.tar.gz ├── meta │ ├── navsim_emu_vla_256_144_trainval_pre_1s_w_token.pkl │ ├── navsim_emu_vla_256_144_test_pre_1s_w_token.pkl │ ├── navsim_emu_vla_256_144_trainval_pre_1s_w_navsim_lat_4s_size_128_num_8_w_bev_reg.pkl │ └── navsim_emu_vla_256_144_test_pre_1s_w_navsim_lat_4s_size_128_num_8_w_bev_reg.pkl └── models ├── Emu3-Stage1.tar.gz ├── fast.tar.gz ├── train_nuplan_6va_v0.2_multi_node.tar.gz ├── train_navsim_lat_200k_size_128_num_8_w_bev_reg.tar.gz └── train_navsim_base_2dyna_2s_size_128_num_8_w_bev_reg_w_dual_query.tar.gz ``` ## Recommended Placement In DynVLA The DynVLA training scripts assume the following runtime layout by default: ```text /the/path/to/DynamicsVLA ├── data │ ├── navsim │ │ ├── BEV_maps │ │ ├── metric_cache │ │ └── processed_data/meta │ └── nuplan │ ├── images │ └── processed_data/vq_codes_low_res_corrected_merge ├── pretrained_models │ ├── Emu3-Stage1 │ ├── fast │ └── train_nuplan_6va_v0.2_multi_node └── logs ├── train_navsim_lat_200k_size_128_num_8_w_bev_reg └── train_navsim_base_2dyna_2s_size_128_num_8_w_bev_reg_w_dual_query ``` File mapping: - `meta/*.pkl` -> `data/navsim/processed_data/meta/` - `BEV/BEV_maps.tar.gz` -> extract to `data/navsim/` - `image_vq/vq_codes_low_res_corrected_merge.tar.gz` -> extract to `data/nuplan/processed_data/` - `image/nuplan_images_256x144.tar.gz.part-*` -> merge and extract to `data/nuplan/`, then rename `nuplan_images_256x144/` to `images/` - `models/Emu3-Stage1.tar.gz` -> extract to `pretrained_models/` - `models/fast.tar.gz` -> extract to `pretrained_models/` - `models/train_nuplan_6va_v0.2_multi_node.tar.gz` -> extract to `pretrained_models/` - `models/train_navsim_lat_200k_size_128_num_8_w_bev_reg.tar.gz` -> optional, extract to `logs/` - `models/train_navsim_base_2dyna_2s_size_128_num_8_w_bev_reg_w_dual_query.tar.gz` -> optional, extract to `logs/` ## Example Setup Commands ```bash export CODE_DIR=/the/path/to/DynamicsVLA export DATA_DIR=/the/path/to/DynamicsVLA export SOURCE_DATA_ROOT=/the/path/to/DynVLA_Training_Data mkdir -p ${DATA_DIR}/data/navsim/processed_data/meta mkdir -p ${DATA_DIR}/data/navsim mkdir -p ${DATA_DIR}/data/nuplan/processed_data mkdir -p ${DATA_DIR}/pretrained_models mkdir -p ${CODE_DIR}/logs cp ${SOURCE_DATA_ROOT}/meta/*.pkl ${DATA_DIR}/data/navsim/processed_data/meta/ tar -xzf ${SOURCE_DATA_ROOT}/BEV/BEV_maps.tar.gz -C ${DATA_DIR}/data/navsim tar -xzf ${SOURCE_DATA_ROOT}/image_vq/vq_codes_low_res_corrected_merge.tar.gz -C ${DATA_DIR}/data/nuplan/processed_data cat ${SOURCE_DATA_ROOT}/image/nuplan_images_256x144.tar.gz.part-* > ${DATA_DIR}/data/nuplan/nuplan_images_256x144.tar.gz tar -xzf ${DATA_DIR}/data/nuplan/nuplan_images_256x144.tar.gz -C ${DATA_DIR}/data/nuplan mv ${DATA_DIR}/data/nuplan/nuplan_images_256x144 ${DATA_DIR}/data/nuplan/images tar -xzf ${SOURCE_DATA_ROOT}/models/Emu3-Stage1.tar.gz -C ${DATA_DIR}/pretrained_models tar -xzf ${SOURCE_DATA_ROOT}/models/fast.tar.gz -C ${DATA_DIR}/pretrained_models tar -xzf ${SOURCE_DATA_ROOT}/models/train_nuplan_6va_v0.2_multi_node.tar.gz -C ${DATA_DIR}/pretrained_models tar -xzf ${SOURCE_DATA_ROOT}/models/train_navsim_lat_200k_size_128_num_8_w_bev_reg.tar.gz -C ${CODE_DIR}/logs tar -xzf ${SOURCE_DATA_ROOT}/models/train_navsim_base_2dyna_2s_size_128_num_8_w_bev_reg_w_dual_query.tar.gz -C ${CODE_DIR}/logs ``` ## Checkpoint Usage - `train_nuplan_6va_v0.2_multi_node.tar.gz` is the released DriveVLA-W0 checkpoint used to initialize DynVLA Stage 1 and Stage 2 training. - `train_navsim_lat_200k_size_128_num_8_w_bev_reg.tar.gz` is the released Stage 1 checkpoint. - `train_navsim_base_2dyna_2s_size_128_num_8_w_bev_reg_w_dual_query.tar.gz` is the released Stage 2 SFT checkpoint. - The released Stage 2 checkpoint can be used directly for Navsim inference or as the initialization checkpoint for DynVLA Stage 3 GRPO training. ## Notes - The default `CODE_DIR` and `DATA_DIR` are the same in the released scripts. - NAVSIM metric cache is not stored in this dataset repository. Please generate it under `data/navsim/metric_cache` by following the main DynVLA README. - Large files are tracked with Git LFS. ## Related Links - DynVLA code repository: https://github.com/yaoyao-jpg/DynamicsVLA - ModelScope dataset page: https://modelscope.cn/datasets/ShuyaoShang/DynVLA_Training_Data