Recently, a research team from Shanghai Jiao Tong University and Shanghai Artificial Intelligence Laboratory released the latest research results "MLM: Learning Multi task Loco Manipulation". This study achieved full body collaborative control of quadruped robots equipped with robotic arms to complete multi task "movement+operation" with a single strategy in the real world, and can be combined with diffusion large models to achieve autonomous operation. Behind this breakthrough, the real-world multitasking trajectory data provided by FastUMI played a crucial role.

MLM Framework: Unified Strategy for Mobile and Operational Collaboration

MLM is a full body reinforcement learning control framework based on real trajectory data and large-scale simulation training. This framework enables quadruped robots equipped with robotic arms to synchronously complete movements and fine operations through a unified strategy without additional real robot training, and to perform various household tasks such as cleaning desktops, pouring drinks, opening containers, grabbing cups, plugging and unplugging chargers, etc.

The research team extracted over 1200 high-quality, multi scenario real operation trajectories from the FastUMI system, covering various typical tasks and providing key data support for MLM training. These trajectories have the characteristics of hardware independence and end-to-end availability, effectively supporting multi task strategy learning in simulation environments and significantly reducing the dependence and cost of real robot training.

Why FastUMI? The value of real-world trajectories is being proven

As an upgrade solution for Universal Operating Interface (UMI), FastUMI provides an efficient solution for robot data collection with its core advantages of "low cost, high adaptability, and easy deployment". This technology reduces the dependence on specific robot hardware through hardware decoupling design, simplifies the system integration process while ensuring visual consistency, and can achieve millimeter level accuracy in typical operation tasks. The collection efficiency is significantly improved compared to traditional methods.  

FastUMI has been validated as a key data source for multi task robot learning in multiple cutting-edge studies, highlighting the irreplaceable value of real-world operation trajectories in training highly generalized and robust robot strategies.

Luming Robotics launches FastUMI Pro: creating an efficient ontology free data acquisition system

Based on the accumulation of FastUMI technology, Luming Robotics has officially launched the FastUMI Pro multimodal ontology free data acquisition software and hardware system. The system adopts a lightweight handheld gripper solution without a robot body, with an overall weight of only 600g and a load capacity of 2kg. It performs well in portability, stability, and continuous operation ability, and supports flexible and efficient data collection in various scenarios.

FastUMI Pro significantly reduces the system complexity and overall cost of data collection, increasing collection efficiency to three times that of traditional solutions and compressing costs to one-fifth. The system has high availability every time it collects data, truly achieving the goal of "high-quality data easy to obtain" and providing a reliable data foundation for robot learning research and industrial applications.

Previously, Luming Robotics has relied on FastUMI technology to accumulate over 10000 hours of high-quality real operation data and conducted base model training. The company will continue to focus on optimizing data collection hardware, software platforms, and models, building a complete technical chain covering data generation, processing, and application, and is committed to promoting the construction of an open and efficient embodied intelligent data ecosystem.Editor/Bian Wenjun