positioned directly in the middle of the net and placed onto the seabed. The system's performance was evaluated by verifying whether the equipment was capable of completing the task. As is shown in Figure 6, the following steps were carried out and evaluated. 1) Approaching the object: The system should move roughly 2 m until the object is within the arm's workspace. 2) Grasping the object: The arm should retrieve the object from the seabed and firmly hold it against hydrodynamic disturbances. 3) Releasing the object: The gripper should be able to open its fingers to allow the object to fall freely. 4) Collecting the object: The target object should be stored in the side-mounted basket. 1) SILVER2 was placed in front of the box, with the manipu- lator fully contracted, and it extended its legs to ensure that the manipulator was above the box. 2) SILVER2 lowered its legs, and the manipulator was able to reach into the box. 3) The manipulator grasped the object via the arm inverse kinematics model and took the article out of the box. The performance during this test was evaluated by measuring the length of the opening area and verifying whether the sys- tem successfully grasped and collected the object. The second test was carried out to collect objects from a simulated overhang environment [Figure 5(e)]. A transparent plate was fixed to a frame to create an overhang barrier, and the object was placed on the seabed. The grasping task con- sisted of two steps: 1) SILVER2 was placed in front of the frame, with the soft manipulator fully contracted. 2) The manipulator grasped the object using the soft arm (via the inverse kinematics model) and took the object from beneath the overhang. The system's performance was evaluated by measuring the height of the plate and verifying whether the manipulator suc- cessfully grasped and collected the object. Collecting Objects in Confined Spaces In this protocol, we investigated whether our soft manipulator could retrieve objects from confined spaces and whether its light weight and compliance would be beneficial for underwa- ter manipulation. Two tests were carried out to estimate the performance of the manipulator in a confined underwater environment. The first focused on collecting objects from a half-open box [Figure 5(d)]. The object and a real-time under- water camera (transmitting images via cables) were placed on the bottom of the box, and half of the box was then covered by a metal plate. The grasping task was executed in three steps: Results We conducted all experiments in realistic sea conditions, on both cloudy and sunny days without rain. The average depth Approach Target Move Toward Target Camera Collecting Basket t=0s (a) (b) Grasp Target (c) Collect Target Task Completed t = 46 s t = 38 s (d) t = 15 s t = 53 s (e) (f) Figure 6. An example of fragile object collection. (a) The robot starts a few meters away from the object and (b) moves close enough so that the object (c) is within the arm's workspace. (d) The soft manipulator grasps the object via the inverse kinematics model and (e) moves it into (f) the onboard collection basket. Scale bar: 15 cm. DECEMBER 2020 * IEEE ROBOTICS & AUTOMATION MAGAZINE * 19