TidyBot++: Household robot cleans, tidies up and tends plants
Household robots don't have to look humanoid. A simple robotic arm on a mobile platform can suffice. However, it must be able to learn.
The TidyBot++ can not only tidy up, but also water plants.
(Image: Jimmy Wu)
A team of scientists from Stanford University, Priceton University and logistics robotics specialist Dexterity have developed TidyBot++, a hypermobile household robot that can learn to perform new tasks through imitation learning. The robot can also be used to collect training data in order to create new algorithms for imitation learning.
TidyBot++ is relatively simple, as the study "TidyBot++: An Open-Source Holonomic Mobile Manipulator for Robot Learning", published on Arxiv, shows. The mobile part for driving the robot is a holonomic base that uses motorized steering wheels. This technology enables the robot to maneuver very precisely in all directions – similar to an office chair. This is important for a domestic robot to be able to position itself precisely in confined spaces to carry out tasks. This allows TidyBot++ to complete its tasks more efficiently without having to constantly readjust its position.
Videos by heise
A simple frame made of aluminum profiles is attached to the mobile base. The frame is easy to build and inexpensive. Several robot arms and the appropriate sensors can be attached to it. In the first version, it only accommodates a single Kinova robot arm with eight degrees of freedom. The platform also carries the batteries and the control electronics of the teleoperation interface for remote control of the robot.
Learning by imitation
To make it easier to learn new tasks, the robot is controlled intuitively using the movements of a cell phone. A human can thus perform the necessary movements for a task, which the robot can then implement by collecting training data and then learning by imitation.
This is where the scientists see the greatest potential for TidyBot++. They believe that the robot can help to collect high-quality training data in order to train AI algorithms for robotics through imitation learning. This does not have to be limited to household tasks, but can also include industrial tasks.
In practice, the scientists have initially limited themselves to training TidyBot++ to work in the household. For example, the robot can already load a dishwasher, wipe tables and take out the garbage. The robot should also be able to water plants.
Further research
In future, its capabilities are to be expanded as part of further studies. The researchers also want to expand the robot's hardware. For example, further sensors for environment recognition and additional robot arms that can work together are to be installed. With several arms, the robot could also greatly expand its range of tasks, for example for tasks that require at least two arms. However, this would make the robot more expensive. So far, its manufacturing costs amount to less than 6000 US dollars.
Other planned research approaches include an automated process to collect large data sets for learning complex manipulation strategies and a way of allowing several robots to work collectively.
(olb)
