New AI Training Trick Helps Robots Learn New Tasks Faster

Imagine teaching a robot to grasp a cup by its handle. You show it a few examples, and it learns. But what if you want it to grasp a different object in a similar way? Normally, you'd have to provide new examples with detailed instructions for each new object. That's time-consuming and expensive. Now, researchers have found a way for robots to learn these preferences without extra human labeling.

The team started with a robot that had been trained on a set of tasks with preference labels—like 'grasp here' or 'move this way.' They found that this robot could follow similar preferences on a new task, but it often failed to complete the task. When they fine-tuned the robot on the new task without labels, it got better at the task but forgot the preferences. To solve this, they created a method called Self-Labeled Preference Transfer (SPT). It uses a shared vision-language model to both recognize preferences and predict actions. After training on labeled examples, the preference recognizer is frozen and used to automatically label new demonstrations, which then train the robot to follow preferences while succeeding at the task.

In simulated tests with five types of preferences, the recognizer was 96.4% accurate. The SPT method achieved 84.8% preference following and 89.2% task success, compared to 59.4% and 88.0% for standard fine-tuning. When multiple preferences were involved, SPT reached 93.8% preference following versus 42.5% for baselines. In real-world experiments, SPT improved preference following from 52.5% to 92.5% across two task transfers.

This means robots can learn new tasks more efficiently, without the need for additional manual annotations. That could make robots cheaper to train and more adaptable to new situations, bringing us closer to robots that can assist in homes, hospitals, and factories with minimal human intervention.