One of the primary use cases in the SMARTHANDLE project involves the handling of aluminium profiles. To facilitate this, KUKA has developed a testbed designed to advance and evaluate technologies related to motion planning, motion execution, object localization, 6D pose estimation, grasp sampling, and grasp evaluation. The testbed is equipped with two robots, each featuring a three-finger gripper, and a camera. Aluminium profiles are manipulated using either single-arm or dual-arm grasps.
The application has the following steps:
- Detection of aluminium profile
A stereo camera generates a point cloud and corresponding 2D images of the current scene. This data is processed by an AI algorithm (CADMatch from Roboception), which determines the 6D pose based on templates derived from CAD data. To enhance the accuracy of the point cloud, particularly for reflective objects, an additional projector is utilised, resulting in improved 6D pose estimations.
- Grasp selection
The grasp selection process comprises two phases. Initially, hundreds of single-arm and dual-arm grasps are generated for all known aluminium profiles and stored in a database. During application runtime, these grasps are evaluated. A score is generated based on the current 6D pose of the object and the corresponding point cloud. The score is influenced by the grasp direction and the collision state, with grasps that are in collision being filtered out.
- Pick and place operation
A collision-free robot trajectory is executed to a pre-pick pose. Upon reaching this pose, an approach motion (currently a linear motion) is performed to grasp the aluminium profile. The profile is then lifted and placed at a designated second pose. All robot motions are planned using novel algorithms that leverage robot redundancies to identify optimal trajectories. Additionally, impedance control is activated for all motions to prevent damage to the aluminium profiles, the robots, and the surrounding environment. This is particularly crucial when handling an aluminium profile with a dual-arm grasp, as small tolerances between the two robots can generate significant forces on the grippers and the profile.
The testbed for aluminium profile handling was successfully established, and initial evaluations of profile picking and placing were conducted. This process involved detecting the aluminium profiles, selecting appropriate grasp points, planning motion sequences for profile handling, and executing these motions using a dual-arm robotic system under impedance control.
Moving forward, the 6D pose estimation will be enhanced to deliver more precise 6D poses for aluminium profiles. In addition, AI-driven grasp planning will be integrated and evaluated to ensure more reliable grasping. Furthermore, the current pick-and-place motion, which currently follows linear trajectories, will be replaced by object-based path planning, specifically designed to adapt to the dynamic requirements of the application environment during runtime.