JTree Labs
We are an open robotics research organization working on full-stack robotic systems research and learning.
We are an open robotics research organization that works on full-stack robotic systems research and learning.
Learn more about our work below:
- Research
- Open Source Projects
- Blog Posts
- Why We Exist
We do our work in San Francisco and El Segundo, CA.
See our latest release
Multitask Diffusion Transformer Policy
Open-Source Release
This method combines the highly scalable Diffusion Transformer (DiT) architecture with the expressive Diffusion Policy representation for robot manipulation. A similar method was demonstrated on the Boston Dynamics Atlas Humanoid robot, where we saw no open-source release and significant appetite from the research community to understand this method.
This inspired us to research it, estimate the architecture from what literature is available, and build an implementation from scratch. This model achieved a high level of performance on a variety of manipulation tasks with only 10–20 hours of training on an H200.
We partnered with the HuggingFace team to release the model as an open-source policy integrated into the LeRobot project. You can find more details about the model, training, and how to use it in the links below.
Blog
It's Not About Modality, It's About Scale
Modern robot learning, like all deep learning, is a numbers game. The question is: how do we approach adding new modalities to robot foundation models from an angle that attacks the bottlenecks that prevent creating and leveraging scale.
Methods for Conditioning Diffusion Models
An overview of different conditioning strategies for diffusion models and their origins in multi-modal generation techniques.
Why Manipulation Is 'Harder' Than Locomotion
Exploring why robotic manipulation presents greater challenges than locomotion for robot learning.