The new training model allowed the AI device to compensate for limb injuries. The robot was able to cover a huge distance without one leg.
The researchers explained that for a device to adapt to a new circumstance, its “brain” needs to be trained in a certain way. Artificial intelligence (AI) often relies on neural networks, algorithms inspired by the human brain. But unlike our organ, the AI brain usually does not learn new actions after graduation.
So in the new study, the researchers embedded Hebb’s rules – mathematical formulas that allow AIs to continue learning – into the network. Instead of values that dictate how activity propagates from one imitation neuron to another, these values change with experience.
To test how their method worked, the team partially removed the robot’s left front leg, forcing it to compensate for the injury on the fly. The device was able to travel seven times the distance of a conventional robot. The researchers reported this at a conference on neuro-information processing systems. Such training can improve algorithms for image recognition, language translation, or driving.
Previously, researchers at MIT created an algorithm that can determine goals and plans, even if they might fail. This type of exploration will improve assistive technology, collaboration or grooming robots, and digital assistants like Siri and Alexa.