# DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation [Haozhe Xie](https://haozhexie.com), [Beichen Wen](https://wenbc21.github.io/), Jiarui Zheng, [Zhaoxi Chen](https://frozenburning.github.io/), [Fangzhou Hong](https://hongfz16.github.io/), [Haiwen Diao](https://paranioar.github.io/), [Ziwei Liu](https://liuziwei7.github.io/) S-Lab, Nanyang Technological University [](https://www.codefactor.io/repository/github/hzxie/DynamicVLA)  [](https://arxiv.org/abs/2601.22153) [](https://youtu.be/NmJnHcI04_Q) [](https://huggingface.co/datasets/hzxie/DOM)  ## Changelog🔥 - [2026/04/26] Released training and testing code. - [2026/01/26] Repository created. ## Cite this work📝 ``` @inproceedings{xie2026dynamicvla, title = {{DynamicVLA:} A Vision-Language-Action Model for Dynamic Object Manipulation}, author = {Xie, Haozhe and Wen, Beichen and Zheng, Jiarui and Chen, Zhaoxi and Hong, Fangzhou and Diao, Haiwen and Liu, Ziwei}, booktitle = {NeurIPS}, year = {2026} } ``` ## Dataset and Pretrained Models 🛢️ ### DOM Dataset - [DOM Training Set](https://huggingface.co/datasets/hzxie/DOM) – for training DynamicVLA - [DOM Testing Set](https://gateway.infinitescript.com/?f=DOM-Test) – for benchmarking; includes test configurations and a subset of 3D scenes - [DOM 3D Objects](https://gateway.infinitescript.com/?f=DOM-3D-Objects) – assets for data generation and benchmarking - [DOM 3D Scenes](https://gateway.infinitescript.com/?f=DOM-3D-Scenes) – full scene assets for data generation ### Pretrained Models - [DynamicVLA (trained on DOM)](https://huggingface.co/hzxie/dynamic-vla-DOM) ## Installation 📥 We recommend using **conda** to create two separate environments: - one for **model training & inference** - one for **Isaac Lab simulation & evaluation** ### PyTorch Environment - Install **Python 3.10** and **PyTorch 2.7.1** *(Other versions should work, but are not fully tested)* - Install dependencies: ```bash pip install -r requirements.txt ``` ### Isaac Lab Environment - Install **Python 3.11** *(Other versions should work, but are not fully tested)* - Install **Isaac Sim 5.1.0** and **Isaac Lab 2.3.2** Follow the official guide: https://isaac-sim.github.io/IsaacLab/v2.3.2/source/setup/installation/index.html - Install additional dependencies: ```bash pip install shapely pyzmq h5py ``` ## Benchmarking Your Policy 🏅 ### Prepare scenes and objects Download [DOM Testing Set](#dom-dataset) (including a subset of DOM 3D scenes) and [DOM 3D Objects](#dom-dataset). ``` PROJECT_ROOT/ ├── objects/ # Put DOM 3D Objects here ├── scenes/ # Put DOM 3D Scenes here | └── textures # Put the textures of 3D scenes here | └── *.usd # Put the USD files of 3D scenes here ├── tests/ # Put DOM Testing Set here | └── *.json |── test-envs.txt # The list of test environments (included in DOM Testing Set) ├── datasets/ # Generated simulated datasets will be stored here └── dynamic-vla/ # git clone https://github.com/hzxie/DynamicVLA dynamic-vla └── runs # Create folder for evaluation and checkpoints output ``` ### Run Policy Evaluation Server > ⚠️ This step requires the **Isaac Lab environment** From the `PROJECT_ROOT/dynamic-vla` directory, run: ```bash python3 simulations/evaluate.py \ --scene_dir ../scenes \ --output_dir ../output/evaluation \ --env_cfg ../test-envs.txt \ --enable_cameras --headless -n 20 --save ``` **Arguments:** - `test-envs.txt` are provided by [DOM Testing Set](#dom-dataset) - `-n 20`: run 20 trials per environment - `--save`: save evaluation videos to `output_dir` - `--headless`: run without GUI - `--enable_cameras`: enable visual observations ### Run Policy Inference > ⚠️ This step requires the **PyTorch environment** From the `PROJECT_ROOT/dynamic-vla` directory, run: ```bash python3 scripts/inference.py \ -p /path/to/vla-checkpoint \ -r euler -d -s ``` **Arguments:** - `-p`: path to the trained model checkpoint - `-r euler`: use Euler angles for rotation representation - `-d`: enable **delta actions** *(actions are relative to current state)* - `-s`: enable **contiguous inference** *(if supported by the model)* ## Simulated Dataset Generation 🧪 > ⚠️ This step requires the **Isaac Lab environment** ### IsaacSim Simulation From the `PROJECT_ROOT/dynamic-vla` directory, you can generate synthetic data using: ```bash python3 simulations/simulate.py \ --headless --enable_cameras --seed 42 --save --task place ``` Example configuration file: `simulations/configs/sim_cfg.yaml` **Arguments:** - `--task` : task type to simulate. Options: `pick`, `place`, `long-horizon`. - `--robot`: robot type used in simulation _(default: `franka`, also supports `piper`)_. - `--headless`: run simulation without GUI. - `--enable_cameras`: include visual observations in the output dataset. - `--debug`: enable debug mode and render trajectories as `.mp4` videos. - `--seed`: random seed for simulation _(automatically increments for each run if specified)_. - `-n`, `--n_simulations`: number of simulation episodes to generate *(default: `10,000`)*. - `--save`: save generated simulation data in HDF5 format. ### Trajectory Replay After data generation, convert the trajectories into a format compatible with VLA training: ```bash python3 scripts/translate_dataset_seq.py \ --dataset_dir ../datasets --output_dir ../datasets-tr \ --enable_cameras --headless --save ``` **Arguments:** - `--dataset_dir`: directory containing the raw simulation datasets. - `--output_dir`: directory to store the processed trajectories. - `--enable_cameras`: include visual observations in the output dataset. - `--headless`: run simulation without GUI. - `--save`: save generated simulation data in HDF5 format. ### Convert LeRobot Dataset We provide a script to convert the generated `.h5` files into the LeRobot dataset *(v2.1 format)*, using **Euler angles** as the rotation representation: ```bash python3 scripts/create_lerobot_dataset.py \ --dataset_dir ../datasets-tr --repo hzxie/DOM --rotation euler ``` This will create lerobot dataset using all the hdf5 datasets in the default output directory. ## Training 👩🏽‍💻 > ⚠️ This step requires the **PyTorch environment** From the `PROJECT_ROOT/dynamic-vla` directory, run: ```bash torchrun --nnodes=1 --nproc_per_node=8 --standalone run.py \ -c configs/dynamicvla.yaml \ -p /path/to/pretrained/model -d hzxie/DOM ``` **Arguments:** - `--nnodes`: number of compute nodes (machines) used for distributed training - `--nproc_per_node`: number of GPUs per node - `-c`: path to the training config file - `-p`: path to the pretrained model checkpoint *(optional)* - `-d`: name of the LeRobot dataset *(v2.1 format)* ### Checkpoint Evaluation > ⚠️ This step requires the **PyTorch environment** During training, you can automatically evaluate checkpoints using the following script: ```bash python3 scripts/eval_checkpoints.py \ -r euler -d -s -p "*fvit*46k*" \ --ckpt_dir ./runs/checkpoints/ \ ``` **Arguments:** - `--ckpt_dir`: directory containing the checkpoints to be evaluated. - `-p`: pattern used to match checkpoint filenames *(supports wildcard patterns)*. - `--host`: host address of the evaluation server *(default: `127.0.0.1`)*. - `--img_port`: port used for the image stream on the evaluation server *(default: `3186`)*. - `--act_port`: port used for the action stream on the evaluation server *(default: `3188`)*. ## License 🗒️ This project is licensed under [NTU S-Lab License 1.0](https://github.com/hzxie/DynamicVLA/blob/master/LICENSE). Redistribution and use should follow this license.