SwapBot: Cheap adapters transform kitchen utensils into robot tools, which learn to cook tofu and return meat

Rather than a universal clamp, the University of Washington relies on interchangeable tactile adapters. Touch increases 15 to 25 points of success to policies learned by imitation.

Sautéing tofu in a pan requires cutting firmly, brushing delicately and applying with precision, each time with a different tool. Based on this observation, a team from the University of Washington presents SwapBot, a modular end-effector system that gives up the search for a universal gripper. The principle: a fixed coupling interface on the robot side, and inexpensive adapters on the tool side, which convert ordinary household utensils, including articulated ones, into effectors that can be used by several robotic platforms.

Each adapter has small touch sensors, changes autonomously and is controlled via an intuitive teleoperation system used to collect demonstrations. User tests show that SwapBot reduces task time and operator workload. From these demonstrations, the researchers trained policies that could cook tofu, mix a salad or flip meat by imitation.

The addition of touch makes a difference on these high-touch gestures: policies trained with tactile demonstrations succeed 15 to 25 percentage points more often than those that rely only on the camera. The authors see this as a practical and extensible path to more robust robotic manipulation.