Our latest embodied foundation model, DM0.5, is officially OUT! 🤖
We reject using predefined tasks to fake the unknown physical world. DM0.5 is built for the gritty, real-world scenes—and we brought the deployment threshold down to a single consumer-level GPU.⬇️
From cucumbers to Chinese yams, we're continuing to explore precision in robotic manipulation.
In our latest demo, Dexmal's DM0.5 enables precise force control to peel a Chinese yam, removing thin layers of skin without cutting unnecessarily deep into the flesh.
Dexmal MaaS, our embodied AI service platform, just got a major upgrade.
First launched in July, Dexmal MaaS now supports an end-to-end workflow, from model integration and data upload to policy training, cloud inference, and real-world robot deployment.
With DexOS Lite, MaaS API, and three post-training pipelines (SFT, DAgger, RECAP), developers can build, adapt, and continuously improve models for real-world robotic tasks.
Who ordered afternoon tea with a side of robotics? 🍞
Robotic arms toast the bread and spread the jam, then coordinate with Dexmal's APEX to deliver the afternoon treat straight to your desk.
Laundry day just got a little smarter. 👕
With Dexmal's DM0.5 in control, our robot takes on everything from T-shirts to button-downs, handling garments of different fabrics, sizes, and shapes. Each piece goes onto a hanger, then neatly onto the rack.
DM0.5 has topped multiple robotics benchmarks over the past few months, and we're excited to see it join the @LeRobotHF ecosystem!
Our model has been open-source for a while. Check out our profile for more DM0.5 demos, updates, and real-world applications.
New in LeRobot: DM0.5 by @Dexmal_AI 🤖
A 4B VLM with a 680M flow-matching action expert, pretrained on 7 robot embodiments. Right now it's the top open-source model on RoboColiseum.
Fine-tune it on your robot:
--policy.path=lerobot/dm05_base
Docs 👇
https://t.co/7OqgQyLY4W
Meet your new bartender. 🍸
Powered by Dexmal's DM0.5, this robot grasps cups of different sizes, adapts to different shapes, and precisely pours liquids into various containers. What looks effortless takes careful perception and precise motion control.
Cheers to embodied AI!
Moving from W1 to ARX requires a bit of few-shot fine-tuning, since differences in arm length, end-effector pose definitions, and camera placement can affect performance, even though both robots have 6-DOF arms. Also, we do expect to achieve comparable success with a small amount of additional data.
Here’s Dexmal’s DM0.5 folding cardboard boxes autonomously, using visual input and robot state to generate continuous actions. It coordinates both arms as the cardboard changes shape and contact conditions change.
We’ve run the same task on both W1 and ARX, demonstrating bimanual control across different robot platforms.
Turns out, robots can be pretty good at fitting in.
Powered by Dexmal's DM0.5, robots tackle a variety of industrial insertion tasks with precision and adapt to new ones through few-shot fine-tuning.
Smarter assembly starts with adaptability. Insert Anything.
Plug & play? More like plug & precision. 🔌
With Dexmal's DM0.5, robots autonomously plug, unplug, and repeat with sub-0.5mm alignment precision, bringing greater reliability to electronics assembly and connector testing.
Dexbotic 2.0 Feature Spotlight: More Ways to Evaluate VLAs 📊
Dexbotic 2.0 now supports two additional benchmarks, VLA-Arena and RoboDojo.
VLA-Arena tests safety, robustness, generalization, and long-horizon behavior across L0–L2 difficulty levels. RoboDojo tests VLA generalization on complex bimanual manipulation tasks.
With these additions, Dexbotic 2.0 now supports a broader range of widely used VLA benchmarks.
Open sourced: https://t.co/ZH1N8UQ70O
Dexbotic 2.0 Feature Spotlight: Observation History 🕒
In a multi-step task, the current camera view may not show what the robot has already done. To decide what comes next, the model needs context from earlier in the task.
Dexbotic 2.0 now supports observation history for DM0.5, helping it track progress through longer tasks.
Open sourced: https://t.co/ZH1N8UQ70O
Dexbotic 2.0 Feature Spotlight: Faster DM0.5 Inference ⚡
Inference speed is the real test of a VLA in deployment.
Dexbotic 2.0's new optional inference component makes DM0.5 up to ~5× faster, with no change to the model, accuracy or hardware.
Your call: go fast, or stay on the default.
Open sourced: https://t.co/ZH1N8UQ70O
Understanding the structure and behavior of the physical world is a big goal for AI. For robotics, that understanding must also translate into action.
At Dexmal, we’re working across world models, robot policies, training, inference, and real-world execution—connecting each step so robots can learn and act in the physical world.
A thoughtful perspective from @drfeifei on where this field is headed. Congratulations to World Labs and AMD on this next chapter!
@mortenlysgaard Yes. The two handheld-view cameras and the camera mounted on the robot arm are inputs to the model. The fixed overhead camera is a separate camera set up only to record the demo and is not used as input to the model.
What happens when the camera shifts in the middle of a task?
The pixels change. The task doesn’t.
Dexmal DM0.5 remains robust to significant viewpoint shifts, recognizing the current task state and continuing execution without losing track.
A different object on the table shouldn’t become another task on a developer’s to-do list.
Dexmal DM0.5 uses the same model weights to recognize, locate, and grasp objects across different shapes, materials, and positions—even with external disruptions. No model swaps or parameter tweaks required.
That's less repetitive engineering for developers, and faster adaptation to whatever the next scenario throws at it.
Watch DM0.5 in action 👇
Some things are easier said than typed. Some tasks are easier shown than described.
With Dexmal’s DM0.5, a human demonstration becomes a Video Prompt. Native long-horizon memory lets it track the full sequence, align with the demonstration, and execute each step in real time.