Robot with a blindfold: Rhoda AI lets its gripper search by touch
The Californian start-up combines a new tactile sensor with its video AI model. Without any camera image, the robot found the balls it was looking for in a box.
Photo: XELA Robotics / Wikimedia Commons, CC BY-SA 4.0 (cropped and resized)
Video: YouTube · Rhoda AI · youtube.com
Palo Alto robotics start-up Rhoda AI has taught a robot a sense of touch. In an experiment presented on 13 October, the head and wrist cameras were covered; the robot had to find the balls with a textured surface in a box by touch alone. According to the company, it retrieved all three textured balls – including two it had never seen in training, which differed in size, firmness and pattern.
Technically, Rhoda is extending its "Direct Video-Action" model: it now predicts not only future camera images but also future tactile signals. A second model translates these predictions into movements, with fresh sensor data fed in continuously. According to Rhoda, the sensor measures forces from a few millinewtons to more than 40 newtons and picks up vibrations at around 500 hertz – it even registered the touch of a feather.
The video also shows wrong grabs and a renewed search of the box, and parts are heavily sped up. The company gives no success rate across many trials. Rhoda AI was valued at 1.7 billion dollars in March.