Bayesian Grasp Planning

We present a Bayesian framework for grasp planning that takes into account uncertainty in object shape or pose, as well as robot motion error. When trying to grasp objects based on noisy sensor data, a common problem is errors in perception, which cause the geometry or pose of the object to be uncertain. For each hypothesis about the geometry or pose of the object to be grasped, different sets of grasps can be planned. Of the resulting grasps, some are likely to work only if particular hypotheses are true, but some may work on most or even all hypotheses. Likewise, some grasps are easily broken by small errors in robot motion, but other grasps are robust to such errors. Our probabilistic framework takes into account all of these factors while trying to estimate the overall probability of success of each grasp, allowing us to select grasps that are robust to incorrect object recognition as well as motion error due to imperfect robot calibration. We demonstrate our framework while using the PR2 robot to grasp common household objects.