$30,000 Robot Can Cook and Do Housework by Itself - Huxiu.com#
#Omnivore
A team of Chinese researchers at Stanford University has developed a low-cost and user-friendly robot operating system called Mobile ALOHA, which enables robots to autonomously perform various household chores and complex tasks. This system combines the dexterity of the ALOHA system's hands with the mobility of a mobile base, allowing robots to learn and execute complex mobile manipulation tasks. The entire system costs less than $32,000, providing an economical and efficient solution for robot learning and mobile manipulation research.
• 💡 The Mobile ALOHA robot operating system enables robots to autonomously perform various household chores and complex tasks.
• 💡 The low-cost robot operating system costs less than $32,000, providing an economical and efficient solution for robot learning and mobile manipulation research.
• 💡 By using imitation learning to train robots to perform complex mobile manipulation tasks, success rates can be improved by up to 90%.
We usually only see robotic butlers in games and movies, but a team from Stanford University has introduced us to the Mobile ALOHA system, which brings robotic butlers closer to reality.
Mobile ALOHA is not a robot itself, but an operating system. With this system, robots can easily perform various delicate tasks, such as frying mushrooms:
Even complex Chinese dishes like "Scallop Braised Chicken" and "Oyster Sauce Lettuce" can be easily prepared by robots. For example, cracking eggs:
Blanching shrimp:
Stir-frying:
With the help of Mobile ALOHA, robots can also perform various complex tasks, such as washing dishes:
Putting chairs back in place:
Wiping tables:
Putting pots into storage cabinets:
Taking the elevator:
And even high-fiving with humans:
In addition, people can also use Mobile ALOHA to remotely control robots to perform delicate household chores such as cleaning windows and sweeping toilets.
From various animations, it is evident that with the support of the Mobile ALOHA system, robots can smoothly perform various household chores and complex tasks.
So, what exactly is Mobile ALOHA?
According to the research team at Stanford University, Mobile ALOHA is a low-cost mobile manipulation platform that combines the dexterity of the ALOHA system's hands with the mobility of a mobile base. The design goal of this system is to enable robots to perform complex mobile manipulation tasks while maintaining low cost and ease of operation.
Using the data collected by Mobile ALOHA, the researchers conducted supervised behavior cloning and found that joint training with the existing static ALOHA dataset significantly improves the performance of mobile manipulation tasks. Even with only 50 task demonstrations, joint training can increase the success rate by up to 90%, enabling Mobile ALOHA to autonomously perform complex mobile manipulation tasks such as stir-frying shrimp, opening double-door cabinets to store heavy pots, calling elevators, and gently rinsing used pans with the kitchen faucet.
Mobile ALOHA is based on the ALOHA system, which is a low-cost dual-arm manipulation device, and Mobile ALOHA adds a remote control system on top of it. To achieve remote control functionality, the researchers installed it on a wheeled base, giving the robot a human-like mobility. The system also includes two wrist cameras and a top camera to capture visual information during operations. Additionally, the system has onboard power and computing capabilities, allowing it to work continuously for hours without an external power source.
In addition to these basic hardware components, the research team designed a system that connects the operator's body to the robot base. The operator is connected to the base through a harness and controls the robot's movement by pulling the base. This design allows the operator to manipulate the ALOHA's arms with their hands while controlling the robot's movement.
Through these interfaces, the research team collected a large amount of operational data. This data includes the linear and angular velocities of the robot base and the joint positions of the robot's arms. This data was used to train imitation learning algorithms to learn how to perform complex mobile manipulation tasks.
Subsequently, the researchers used supervised behavior cloning to train the robot. They first used the joint positions of the robot and the velocities of the base as action vectors, and then combined these action vectors with the robot's observations to form a 16-dimensional action vector. This approach allows Mobile ALOHA to directly benefit from previous deep imitation learning algorithms with minimal changes to the implementation strategy.
To improve the performance of imitation learning, the researchers adopted a joint training approach. They combined the data collected by Mobile ALOHA with the existing static ALOHA dataset for training. This joint training approach demonstrated positive transfer in almost all mobile manipulation tasks, achieving equivalent or better performance and data efficiency even in cases with different tasks and morphologies.
With these implementation principles, the Mobile ALOHA system can master complex mobile manipulation tasks through imitation learning with limited demonstration data. This low-cost solution provides researchers with a practical platform for studying and developing robots capable of performing practical tasks in a home environment.
It is worth mentioning that the team has also disclosed the cost of the Mobile ALOHA system. The total cost of the entire system is less than $32,000, including robot hardware, power and computing devices, cameras, sensors, assembly and maintenance costs, and the open-source software component.
Mobile ALOHA provides an economical and efficient solution for robot learning and mobile manipulation research, allowing more researchers and developers to participate in this field.
Although Mobile ALOHA has made significant progress in both hardware and software, the team also acknowledges that there are still limitations, such as the system's large footprint and the difficulty for fixed-height arms to reach lower cabinets, ovens, and dishwashers. In the future, their work will focus on addressing these hardware limitations and exploring how to perform imitation learning from highly suboptimal and heterogeneous datasets.
The Mobile ALOHA project is currently open-source on GitHub, and the team has also released corresponding papers and introductions. This technology is not yet mature, and the developers have stated that they will release more detailed papers on the Arxiv platform in the near future. GenAI will continue to follow the technology details and provide timely interpretations of any new information.