hzxie/DynamicVLA

The official implementation of "DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation". (NeurIPS 2026)

What it solves

DynamicVLA is designed to enable robots to manipulate objects that are moving or changing state (dynamic object manipulation), moving beyond the limitations of static scene manipulation.

How it works

It is a Vision-Language-Action (VLA) model that processes visual observations and language instructions to output robot actions. The project provides a complete pipeline for training and evaluation, including a specialized dataset called DOM (Dynamic Object Manipulation) and integration with Isaac Lab for high-fidelity simulation and benchmarking.

Who it’s for

Researchers and developers working on embodied AI, robotic manipulation, and VLA models who need to handle dynamic environments.

Highlights

  • DOM Dataset: A dedicated dataset for training and benchmarking dynamic object manipulation.
  • Simulation Integration: Full support for Isaac Lab and Isaac Sim for generating synthetic data and evaluating policies.
  • Flexible Action Representation: Supports delta actions (relative to current state) and Euler angles for rotation.
  • Distributed Training: Includes scripts for multi-GPU distributed training using torchrun.

Related

  • Project
    InternRobotics/InternVLA-M1

    InternVLA-M1 is a spatially guided vision-language-action framework that enables generalist robot policies to perform complex tasks by combining vision-language understanding with robotic action prediction.

  • Project
    dexmal/opendm

    OpenDM is a Vision-Language-Action (VLA) model and framework for open-world robot control, enabling robots to follow open-ended instructions and handle complex, long-horizon tasks.

  • Project
    InternRobotics/VLAC

    A Vision-Language-Action-Critic model that provides dense reward feedback and trajectory quality screening for real-world robotic reinforcement learning.

  • Project
    Robbyant/lingbot-vla

    A pragmatic Vision-Language-Action (VLA) foundation model pre-trained on 20,000 hours of real-world robot data to enable high-performance robotic control across multiple platforms.

  • Project
    2toinf/X-VLA

    X-VLA is a soft-prompted Transformer-based Vision-Language-Action model designed for scalable cross-embodiment robotic control across diverse hardware platforms.