While robots in many cases can achieve substantial efficiency gains over manual labor, such systems struggle to do many tasks that are simple for humans. Warehouse picking tasks, for example, are difficult for robots because items may be partially or completely hidden from view, preventing the robot from locating them quickly.
A research group at MIT is working to give robots superhuman abilities that may render these issues a problem of the past. Their robotic grasping system, RF-Grasp, can see through obstructions using radio waves to quickly locate and retrieve even totally obstructed objects.
Radio frequency identification (RFID) tags are attached to each object for identification purposes. RF-Grasp then uses an RFID reader to determine the approximate location of an object of interest. Since these radio waves can pass through other objects, visual obstructions will not hinder this initial step. With the approximate location determined, a camera attached to the robot’s wrist is used to zero in on the precise location. If the object is not immediately visible, the robot’s control algorithm moves obstructing objects out of the way to uncover, and then retrieve, the target item.
Combining both radio frequency data and camera input makes for a very complicated motion planning algorithm. To provide a solution to the problem, the team designed a deep-reinforcement learning network, consisting of two 121-layer DenseNets. One network takes camera data as input, while the other accepts radio data. Pretrained DenseNet models were employed, which were then further fine-tuned by training on data generated in a simulated environment.
RF-Grasp was compared with similar robots that were only equipped with cameras. In a series of trials, the new method was found to improve success rate and efficiency by up to 50% over the current state of the art. The more occlusions that are present, the more the performance of RF-Grasp exceeds camera-only solutions, highlighting the value added by RFID sensing.
The researchers are now working on enhancing the abilities of RF-Grasp. Currently, the system is limited in the types of objects it can grasp, and the angles at which it can grasp them. They believe this can be remedied by replacing the present deep-reinforcement learning network with one better suited to more complex setups. Also, as the device uses separate components for RF localization and grasping, there is an initial calibration step required for proper RF-eye-hand coordination. They would like to develop a more fully integrated solution that would make this calibration unnecessary.