Mon Sep 09 16:10:06 UTC 2024: ## Robots Now Learn to Adapt to New Environments Without Training
New York University researchers have developed a revolutionary system called “Robot Utility Models” (RUMs) that enables robots to perform tasks in entirely new environments without any additional training or data. This groundbreaking technology achieves an impressive 90% success rate across various tasks and environments, marking a significant step towards truly intelligent robots.
Previously, robots relied heavily on specific training data for each environment, limiting their ability to adapt to new situations. RUMs, however, overcome this limitation by leveraging multi-modal imitation learning and a powerful language model (LLM) to analyze the robot’s observations and determine whether a task has been completed.
The researchers trained five utility models for tasks like opening cabinets, picking up objects, and reorienting fallen items. These models were tested on a commercially available robot, the Hello Robot Stretch, and demonstrated success in unseen environments with objects the robot had never encountered before.
The team also highlighted the crucial role of high-quality and diverse training data in developing these adaptable models. They emphasized the need for multiple demonstrations of each task in a variety of environments to ensure the robot can generalize its learned skills.
This breakthrough in robotics opens new possibilities for robots to seamlessly operate in dynamic and unpredictable settings. RUMs have the potential to revolutionize various industries, enabling robots to perform complex tasks in diverse environments without the need for constant human intervention.
First Piper 