AI-Powered Robot Arms: Precise Enough to Handle Pringles Chips

The team’s groundbreaking progress lies in the use of two robotic arms, which is a departure from the typical single-arm approach in most tactile robotic projects. Surprisingly, despite the doubling of limbs, the training process only requires a mere few hours.

The researchers achieve this by initially training their AI in a simulated environment and subsequently implementing the fully developed Bi-Touch system on their physical robotic arms.

“With our Bi-Touch system, we can easily train AI agents in a virtual world within a couple of hours to achieve bimanual tasks that are tailored towards the touch,” Lin continued. “And more importantly, we can directly apply these agents from the virtual world to the real world without further training.”

The remarkable success of the Bi-Touch system can be attributed to its reliance on a technique known as Deep Reinforcement Learning (Deep-RL). In this approach, robots learn by repeatedly attempt tasks, similar to trial-and-error experimentation.

When the robot accomplishes a task successfully, researchers reward the AI, much like how you would train a pet with positive reinforcement. Over time, the AI learns the most effective steps to achieve its designated objective. In this case, the goal is to use the two robotic limbs, each equipped with a soft pad, to pick up and manipulate objects like a foam brain mold, a plastic apple, or an individual chip.

Notably, the Bi-Touch system doesn’t rely on visual input; instead, it relies solely on proprioceptive feedback, including factors like force, physical positioning, and self-movement.