Eye-in-Hand Autonomous Exploration and Mapping of a 3D Unknown Workspace via Virtual Model Control on a 7-DOF Manipulator

Various robotic applications require the ability to autonomously explore and map a space, particularly in dangerous environments, where human access is restricted, such as mines or nuclear sites. Nowadays, even harvesting is automated, requiring the ability to create a map of the plantation to keep track of the changes and record them. This thesis proposes a new method for autonomous exploration and mapping of unknown spaces with a 7-dof anthropomorphic arm governed by the physics of virtual mechanical elements connected between the internal model of the robot and the internal model of the workspace (VMC: Virtual Model Control). The output is a 3D grid of voxels that divides the space into occupied and free, based on camera readings. To acquire new information on the objects populating the workspace, a depth camera is mounted on the robot's gripper, returning a point cloud of the acquisition. The VMC continuously pulls and orients the camera toward the unknown voxel with the highest multi-criteria score overall. The composite score rates the voxels according to three different criteria: density of unknown voxels in the surroundings, distance from the camera, distance from the robot's base. Obstacle avoidance is performed by attaching repulsive springs between collision points distributed on the robot and occupied voxels (or the floor). The method has a common limitation of the potential-field-based approaches: stalls. The stalls, or deadlocks, occur in the local minima, where the sum of forces is zero. The algorithm adopts a two-way recovery system that resolves 100% of the stalls and allows a continuous exploration. Numerous real-world and simulation experiments validate the method and its correctness: exploration does not achieve full completeness, due to unreachable poses for the discovery, but stops at a fixed time limit, when over 90% (and more) of the workspaces have been covered.