--- license: mit pipeline_tag: robotics tags: - vision-language-action-model - vision-language-model --- # Model Card for InternVLA-M1_object InternVLA-M1 is an open-source, end-to-end vision–language–action (VLA) framework for building and researching generalist robot policies, as introduced in the paper [InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy](https://huggingface.co/papers/2510.13778). - 🌐 Homepage: [InternVLA-M1 Project Page](https://internrobotics.github.io/internvla-m1.github.io/) - 💻 Codebase: [InternVLA-M1 GitHub Repo](https://github.com/InternRobotics/InternVLA-M1)
## Abstract We introduce InternVLA-M1, a unified framework for spatial grounding and robot control that advances instruction-following robots toward scalable, general-purpose intelligence. Its core idea is spatially guided vision-language-action training, where spatial grounding serves as the critical link between instructions and robot actions. InternVLA-M1 employs a two-stage pipeline: (i) spatial grounding pre-training on over 2.3M spatial reasoning data to determine "where to act" by aligning instructions with visual, embodiment-agnostic positions, and (ii) spatially guided action post-training to decide "how to act" by generating embodiment-aware actions through plug-and-play spatial prompting. This spatially guided training recipe yields consistent gains: InternVLA-M1 outperforms its variant without spatial guidance by +14.6% on SimplerEnv Google Robot, +17% on WidowX, and +4.3% on LIBERO Franka, while demonstrating stronger spatial reasoning capability in box, point, and trace prediction. To further scale instruction following, we built a simulation engine to collect 244K generalizable pick-and-place episodes, enabling a 6.2% average improvement across 200 tasks and 3K+ objects. In real-world clustered pick-and-place, InternVLA-M1 improved by 7.3%, and with synthetic co-training, achieved +20.6% on unseen objects and novel configurations. Moreover, in long-horizon reasoning-intensive scenarios, it surpassed existing works by over 10%. These results highlight spatially guided training as a unifying principle for scalable and resilient generalist robots. ## Training Details ``` action_chunk: 8 batch_size: 128 training_steps: 30k ``` ## Sample Usage Below are two collapsible examples: InternVLA-M1 chat and action prediction.
InternVLA-M1 Chat Demo (image Q&A / Spatial Grounding) ```python from InternVLA.model.framework.M1 import InternVLA_M1 from PIL import Image import requests from io import BytesIO import torch def load_image_from_url(url: str) -> Image.Image: resp = requests.get(url, timeout=15) resp.raise_for_status() img = Image.open(BytesIO(resp.content)).convert("RGB") return img saved_model_path = "/PATH/checkpoints/steps_50000_pytorch_model.pt" internVLA_M1 = InternVLA_M1.from_pretrained(saved_model_path) # Use the raw image link for direct download image_url = "https://raw.githubusercontent.com/InternRobotics/InternVLA-M1/InternVLA-M1/assets/table.jpeg" image = load_image_from_url(image_url) question = "Give the bounding box for the apple." response = internVLA_M1.chat_with_M1(image, question) print(response) ```
InternVLA-M1 Action Prediction Demo (two views) ```python from InternVLA.model.framework.M1 import InternVLA_M1 from PIL import Image import requests from io import BytesIO import torch def load_image_from_url(url: str) -> Image.Image: resp = requests.get(url, timeout=15) resp.raise_for_status() img = Image.open(BytesIO(resp.content)).convert("RGB") return img saved_model_path = "/PATH/checkpoints/steps_50000_pytorch_model.pt" internVLA_M1 = InternVLA_M1.from_pretrained(saved_model_path) image_url = "https://raw.githubusercontent.com/InternRobotics/InternVLA-M1/InternVLA-M1/assets/table.jpeg" view1 = load_image_from_url(image_url) view2 = view1.copy() # Construct input: batch size = 1, two views batch_images = [[view1, view2]] # List[List[PIL.Image]] instructions = ["Pick up the apple and place it on the plate."] if torch.cuda.is_available(): internVLA_M1 = internVLA_M1.to("cuda") pred = internVLA_M1.predict_action( batch_images=batch_images, instructions=instructions, cfg_scale=1.5, use_ddim=True, num_ddim_steps=10, ) normalized_actions = pred["normalized_actions"] # [B, T, action_dim] print(normalized_actions.shape, type(normalized_actions)) ```
## Acknowledgements We thank the open-source community for their inspiring work. This project builds upon and is inspired by the following projects (alphabetical order): - [IPEC-COMMUNITY](https://huggingface.co/IPEC-COMMUNITY): Curated OXE / LIBERO style multi-task datasets and formatting examples. - [Isaac-GR00T](https://github.com/NVIDIA/Isaac-GR00T): Standardized action data loader (GR00T-LeRobot). - [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL/blob/main/qwen-vl-finetune/README.md): Multimodal input/output format, data loader, and pretrained VLM backbone. - [CogACT](https://github.com/microsoft/CogACT/tree/main/action_model): Reference for a DiT-style action head design. - [Llavavla](https://github.com/JinhuiYE/llavavla): Baseline code structure and engineering design references. - [GenManip Simulation Platform](https://github.com/InternRobotics/GenManip): Simulation platform for generalizable pick-and-place based on Isaac Sim. Notes: - If any required attribution or license header is missing, please open an issue and we will correct it promptly. - All third-party resources remain under their original licenses; users should comply with respective terms. --- Thanks for using **InternVLA-M1**! 🌟 If you find it useful, please consider giving us a ⭐ on GitHub. ## Citation ```bibtex @article{internvlam1, title = {InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy}, author = {InternVLA-M1 Contributors}, journal = {arXiv preprint arXiv:2510.13778}, year = {2025}, url = {https://huggingface.co/papers/2510.13778} } ```