United States – A new robotic learning system allows a robot to perform a variety of tasks by first learning from humans and then improving its abilities through additional training.

The system, called RL-100, combines imitation learning, where robots copy human demonstrations, with reinforcement learning that helps the robot refine its movements and become more reliable.

Researchers tested the framework in simulations and in eight real-world tasks involving rigid and deformable objects. The robot successfully completed tasks including folding a towel, unscrewing a lid, rolling a bowling ball, building a box, preparing orange juice, and pouring liquids.

The robot was also able to adapt to new task variations, handle clutter and obstacles, and continue working when faced with human interference.

Using the system, a robot prepared fresh juice for seven hours at a public mall without a break and without encountering any failures.

Researchers compared the robot’s performance with human teleoperators and found that RL-100 matched and often exceeded human speed and quality in completing tasks.

“Imitation is not the end of learning but the beginning of it,” Johannes Stork wrote about the system’s approach.

Podcast: Play in new window | Download