Researchers at ETH Zurich have successfully trained the ORCA robotic hand to write with a pen in a remarkable time frame of just 18 seconds, without needing thousands of training samples or human demonstrations. Initially, the robotic hand spends approximately 18 seconds observing how its finger movements influence the pen tip. Following this brief observation period, it can proficiently write letters and create shapes, continuously fine-tuning its movements in real time.
The system operates without pre-calculating the intricate physics involved in the interaction between the pen and fingers. It utilizes a camera to monitor the pen tip, while an algorithm dynamically assesses how the movements of the hand's joints impact the pen.
This writing capability serves as a showcase for the technology. The researchers have proposed their method as a viable alternative to both reinforcement learning and imitation learning, which typically necessitate extensive data collection over hundreds of hours and significantly greater computational resources.
Remarkably, the entire algorithm, encompassing computer vision, was executed on a standard laptop CPU. This innovative approach shows promise for potential applications in controlling prosthetic devices and small robots.
Informational material. 18+.
