MIT develops a new framework to enable robots to quickly learn in new environments

2023-12-19

Recently, MIT researchers have developed a framework that allows non-technical users to help robots quickly adjust and learn skills in new environments. This technology helps household robots overcome potential failures in real environments and achieve smooth collaboration.

The R&D team stated that most existing household robots are pre trained in factories before leaving the factory, but there are differences between the training environment and the actual situation at the user's home, which may lead to robot failures during operation. Previous methods required users to re demonstrate to train robots, but the efficiency was not high.

To address this issue, the MIT team designed an algorithm system that generates "counterfactual" explanations when a robot fails, indicating which elements need to be changed in order to successfully complete the task. The system then presents these explanations to users and solicits additional feedback.

For example, if a robot fails to grab a marked mug on a table, the system will generate counterfactual explanations such as "changing the cup color", and the user can point out that the cup color is irrelevant to completing the task.

Through this, the system can identify which visual elements are not important to the task, and then generate new data through data augmentation to retrain the robot, thereby achieving rapid generalization of similar tasks.

Compared with traditional methods that require a large number of user demonstrations, this method requires fewer training samples and has better learning performance.

The MIT team stated that this framework allows ordinary users to participate in the training process of robots, promoting better adaptation of household robots to real-life environments. This research achievement demonstrates a new direction of collaboration between robots and humans, and is expected to accelerate the popularization and application of robot technology.

Original title: MIT develops a new framework that enables robots to learn quickly in new environments without the need for professional knowledge

Analysis of the application of ammonia in the refrigeration system of fishing boats