Center-of-Mass Location Estimation for Improving the Adaptability of Grasping Failure Recovery Policy

Vision-based grasping policies may fail on unknown center-of-mass objects (UCMOs) due to the inaccurate estimation of center of mass (CoM). While force and tactile sensors, as well as reinforcement learning (RL) methods, have been utilized to recover from grasping failures, the adaptability is limited to single-axis and simple multiaxis objects (objects with multiple branches). To enhance the adaptability of the grasping failure recovery policy for UCMOs, we propose explicitly estimating the CoM using a novel Bayesian method. Additionally, we introduce a statistical description of task performance and derive a unified metric for evaluating the reliability and efficiency of grasping policies. Compared to the state-of-the-art RL-based method, our approach demonstrates similar efficiency and higher reliability on single-axis objects, with an average success rate of 93.6% and an average number of attempts of 4.39. Moreover, on complex multiaxis objects, our method outperforms the RL-based method (92.0% versus 64.0% and 3.96 versus 6.36). The proposed method improves the adaptability of the grasping failure recovery policy and requires no pretraining.