Google DeepMind: How new AI systems enable robots to tie shoelaces
Googe DeepMInd has launched two new AI systems that can also be used to teach two-armed robots complex tasks.
(Image: Google DeepMind)
Google's AI forge DeepMind has developed two new artificial intelligence (AI) systems, ALOHA Unleashed and DemoStart, to help robots learn complex tasks that require special dexterity with two arms. ALOHA Unleashed enables robots to learn two-armed manipulation tasks. DemoStart uses simulations to improve the performance of a multi-fingered robotic hand in the real world.
Manipulation of objects is possible for most AI-controlled robotic systems with only one arm. Google DeepMind's new AI systems aim to change that. ALOHA Unleashed is based on the ALOHA-2 platform, which consists of an open-source robotic system with two arms and multi-fingered hands that are designed to be more dexterous than those of many other systems. The arms and hands are controlled by an operator who generates training data. This data can be used to teach the robot new tasks. The advantage of this method is that it requires fewer demonstrations to learn complex tasks.
The researchers at Google DeepMind controlled the robotic system in such a way that the two arms tied the laces of a shoe and hung a T-shirt on a coat hanger. The scientists used a diffusion method to do this. This allows the robot's actions to be predicted from random noise. This allows the robot to learn from the data in such a way that it then performs the same tasks independently.
Learning in the simulation with DemoStart
Wherever the task becomes even more complex, such as object manipulation with multi-fingered robot hands, DemoStart comes into play. This is essentially a simulation that works with reinforcement learning algorithms. Initially, simple states are used for learning, with increasingly difficult states being added over time. This continues until the robot performs a task in the best possible way. The system should require around 100 times fewer simulated demonstrations than is the case when learning from real examples.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmmung wird hier ein externes Video (Kaltura Inc.) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Kaltura Inc.) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
In the simulation with DemoStart, the robot achieved a success rate of more than 98 percent for various tasks, such as aligning colored cubes, tightening nuts and cleaning up tools. In the real world, the robot system trained in this way achieved a success rate of 97 percent when aligning the cubes and a success rate of 64 percent when inserting plugs into a socket. According to Google DeepMind, the latter requires significantly greater finger coordination and precision. The success rate is therefore lower.
The researchers at DeepMind see the combination of generating training data with a few demonstrations using ALOHA Unleashed and reinforcement learning with DemoStart as a way of effectively bridging the gap between simulation and reality. This makes it easier to transfer simulation data to a physical robot and reduce the costs and time required.
(olb)