Target Grasp-conditioned Whole-body Control of Satellite Manipulators using Neural Fields with MPC

This paper presents a method for performing whole-body control of orbiting space robots that include grasp planning for proximity control operations while accounting for the diversity in possible geometries of target bodies. Proximity control operations aimed at capturing satellite debris using robotic manipulator arms mounted onboard chaser satellites in orbit are executed in multiple phases. Trajectory planning and control of the satellite-arm coupled system while trying to reach the target constitutes the approaching phase of the proximity operations. This phase presents two major challenges. Firstly, the satellite base is floating, leading to deviations in the base caused by the motion of the arm, which in turn disturbs the trajectory of the arm. Secondly, as the geometry of the target debris is not necessarily well-known a priori, the optimal grasp pose on the target is uncertain. Identifying the most feasible grasp point on the target object is a major challenge that remains to be solved in order for the manipulator to approach the target along an optimal path. We have, in our previous work, developed Model Predictive controllers (MPC) aimed at solving the first challenge of coupled control of space robots while reaching the target object. However, the grasp point was assumed to be known, an assumption that would not hold for non-cooperative operations like debris capture or servicing an old satellite. In this paper, we first estimate the most feasible grasp points on the target body. To achieve this, we incorporate a neural network-based Grasp Diffusion Field (GDF) into the whole-body MPC framework of the space robot. The input to the GDF is a partial point cloud of the target body obtained from a depth camera onboard the satellite, and the output is a set of candidate 6-D grasp poses on the object. This is compared with a convolutional neural network-based method for grasp generation using point clouds. For each method, we use the grasp candidate with the highest score as the target for MPC. We evaluate these methods in PyBullet simulations and perform proximity control operations. Results show that the GDF-based approach is able to find suitable grasp poses on previously unseen, irregularly shaped target objects, leading to successful proximity control operations.