EleTac elephant-trunk-inspired gripper

[00:00:00]Introducing L-Attack, a soft grip inspired by the tip of an elephant's trunk, with visual-tactile perception and proprioception. An elephant's trunk can manipulate objects of any shape and size, from small pieces of cheese to this flat and wide tortilla. It is also one of the most touch- sensitive body parts of all animals.

[00:00:24]Inspired by the tip of an elephant's trunk, we designed the L-Attack grip, capable of diverse gripping and tactile sensations; The L-Attack body is a thin hollow shell. When creating negative air pressure, two of his fingers perform a pinching motion.

[00:00:41]Because the fingers are soft, the grip can hold objects of various shapes and sizes. According to our tests, the L-Attack is best suited for working with light objects weighing up to 30 grams. Small objects can also be captured without any difficulty. Thanks to the softness of the grip, fragile objects such as a grape or even a piece of tofu can be lifted without damaging them. A thin, flat piece of fabric can also be easily pinched and lifted. Interestingly, thanks to its own malleability, L-Attack can pull a card from the deck with simple control commands. Regarding perception, L-Attack is capable of performing several functions. The first is the assessment of the contact.

[00:01:42]Using tactile images captured by its camera, L-Attack can detect contact on both fingers and predict the location and strength of the contact force. Secondly, L-Attack can determine the geometric shape of the objects it touches. In this study, he can recognize six shapes: circle, ring, square, stripe, dots, and triangle.

[00:02:11]In the video on the left, we test it on 3D-printed samples of these six well-known geometric shapes.

[00:02:17]Whereas in the video on the right, he interacts with random objects, such as a bolt or a 3D-printed toy fish.

[00:02:33]Finally, L-Attack is capable of proprioception, assessing the degree to which its fingers are opened or closed. For this feature, we use a deep learning network trained on simulation data.

[00:02:45]During operation, real images are converted into simulated images using a generative model.

[00:03:00]We demonstrate this feature in dynamic scenarios. In the first experiment, L-Attack tracks the movement of the gripper panel, changing its own height and width depending on the gripper movement. Next, L-Attack determines the direction of inclination of the handle slot, which you need to know before inserting the handle. The gripper lowers the handle to the edge of the slot while the robot's wrist swings from side to side. The recorded signals show clear patterns for the two tilt directions, confirming that the proprioceptive function provides useful information about the environment.

[00:03:37]By determining the direction of inclination of the pin socket, we can control the robotic arm to align the handle with the socket, ensuring smoother insertion. We demonstrate the L-attack in two practical applications.

[00:03:58]First, the grab finds and digs out an object buried in the granular material. When he touches the hidden handle, the sensor signals change.

[00:04:08]By searching the grid, we determine the location and orientation of the handle, and then align the grip to grab and lift it.

[00:04:18]Finally, we demonstrate the L-attack while holding a cutlery cleaning sponge. As the sponge moves over uneven surfaces, proprioceptive feedback changes and is used to adjust the movement of the robotic arm.

[00:04:31]The effectiveness is obvious as the red ink on the cutlery is removed. Thank you for watching.