Moving beyond the traditional paradigm of adapting internet-pretrained models to physical tasks, we present DM0, an Embodied-Native Vision-Language-Action (VLA) framework designed for Physical AI. Unlike approaches that treat physical grounding as a fine-tuning afterthought, DM0 unifies embodied manipulation and navigation by learning from heterogeneous data sources from the onset. Our methodology follows a comprehensive three-stage pipeline: Pretraining, Mid-Training, and Post-Training. First, we conduct large-scale unified pretraining on the Vision-Language Model (VLM) using diverse corpora--seamlessly integrating web text, autonomous driving scenarios, and embodied interaction logs-to jointly acquire semantic knowledge and physical priors. Subsequently, we build a flow-matching action expert atop the VLM. To reconcile high-level reasoning with low-level control, DM0 employs a hybrid training strategy: for embodied data, gradients from the action expert are not backpropagated to the VLM to preserve generalized representations, while the VLM remains trainable on non-embodied data. Furthermore, we introduce an Embodied Spatial Scaffolding strategy to construct spatial Chain-of-Thought (CoT) reasoning, effectively constraining the action solution space. Experiments on the RoboChallenge benchmark demonstrate that DM0 achieves state-of-the-art performance in both Specialist and Generalist settings on Table30.
Figures & tables
Figure 2: Model Architecture. DM0 consists of a vision-language model (VLM) backbone and a Flow Matching ( Lipman et al., 2022 ) based action expert. The VLM processes multi-modal inputs and generates embodied reasoning representations, which are subsequently consumed by the action expert to produce continuous robot actions.
Figure 3: Overview of Curated Vision-Language Data. The curated dataset is designed to enhance core embodied reasoning abilities while preserving the general multimodal understanding and reasoning capabilities of the pretrained VLM.
Figure 4: Data mixture ratios for pretraining, midtraining and post-training. After weighted sampling, the total data volumes for the three stages are 1.13T tokens (pretraining), 200M samples (mid-training), and 50M samples (post-training).
Image
QAs
Q: What type is the rack? A: wooden rack. Q: Q: How many bowls in the image? A: 4. Q: Q: What is the object that the robot arm is grasping? A: Cup.
Q: Describe the image. A: The scene depicts a white table with various items scattered across its surface. The items include a red toothpaste tube, a red box, a black trash can, a clear plastic container, a white box, a black marker, a blue marker, a pink cup, a yellow marker, a red marker, a white marker, a black eraser, and a small white box.
Q: Describe the content of image. A: The image shows a pair of robotic arms positioned on a white table, gripping the handles of a blue fabric. To the left, a gray bag contains red fabric. An orange folded item rests on the table to the right.
Q: What is the color of clothes in the right box? A: clothes in the right box are blue. Q: Q: What is the color of clothes in the middle box? A: clothes in the middle box are black.
Table 3: DM0 retains fundamental VQA capabilities such as scene description, object detection, and attribute recognition in embodied scenarios.
Image
QAs
Q: Describe the image. A: The image shows a dog with a light brown coat, sitting and smiling at the camera. The dog’s eyes are open and its mouth is slightly open, revealing a pink tongue. The background is blurred, but it appears to be a indoor setting with a white wall and a white floor. The dog’s ears are slightly pointed, and its face is clean and well-groomed. The overall mood of the image is cheerful and friendly.
Q: What is the name of this book? A: The name of the book is "Algebra" by Serge Lang. Q: Q: What color is the word "Algebra"? A: The word "Algebra" is black.
Q: Descibe the image. A: The image captures a bustling city street, teeming with life and activity. The street, marked with a crosswalk, serves as the main focus of the image. It’s lined with tall buildings, their glass facades reflecting the overcast sky above. The buildings vary in height and design, creating a dynamic skyline that adds depth to the scene. A few cars are visible on the street, their presence adding a sense of movement and life to the image. People can be seen walking on the sidewalk, going about their day, their figures small against the backdrop of the cityscape. The sky above is filled with clouds, casting a soft light over the scene. Despite the overcast conditions, the image exudes a sense of energy and vibrancy, characteristic of a busy city street.
Table 4: DM0 retains fundamental VQA capabilities such as scene description, OCR in general scenarios.
Image
QAs
Q: Please decompose the specified high-level goal ’organize toys’ into its constituent subtasks. A: Pick up the toy car from the ground with your right hand and place it into the storage box.
Q: Please identify the initial subtask required to achieve the following objective: open and close microwave oven. A: Place the plate from the microwave back on the right side of the table with your right hand.
Q: Please identify the specific subtask required to achieve the following goal: lift the steamer lid. A: Pick up the second steamer lid from the right with your left hand and place it on the third steamer from the right.
Q: In order to complete the instruction ’fold clothes’, please identify the necessary subtask plan. A: The current subtask is determined to be: Fold the red T-shirt on the bed neatly with both hands.
Table 5: DM0 possesses Chain-of-Thought (CoT) capability for predicting subtasks within embodied scenarios.
Image
QAs
Q: I want to order my dinner, which button should I press? A: Press the 外卖 button.
Q: I want to check which games are in discounts, which button should I press? A: To check which games are in discounts, press the "More" button located at the top right corner of the "Discounts & Events" section.
Table 6: DM0 also demonstrates potential for supporting mobile agents.