Efficient Human-like Dexterous Grasping under Uncertainty for Robotics
Imagine that you are reaching into the fridge to grasp an object you can only partially see. Rather than relying solely on vision, you must use touch in order to localise it and securely grasp it. However, humans would not poke the object to localise it first and then grasp it. We compensate for the uncertainty by approaching the object in a way such that if a contact occurs it will generate enough information about where the object is and the object will be grasped with a minimum adaptation of the initial trajectory.
The figures above show the observational model for tactile information. The poses p1 and p2 represent two hypothesised configurations of a mug to be grasped. The dotted lines show two possible trajectories for a finger to reach and touch the mug. Hypothesis p1 represents the expected mean pose for the mug. The left figure shows the expected contact signal for both hypotheses along the trajectory. At time tj the planner expects to observe a contact if the object is in pose p2 and no contact for pose p1. In the right figure, the planner expects similar observations in both cases at time tk. Thus the trajectory on the left is more likely to distinguish hypothesis p1 versus p2 than the trajectory on the right.
Humans would not poke the object to localise it first and then grasp it. We compensate for the uncertainty by approaching the object in a way such that if a contact occurs it will generate enough information about where the object is and the object will be grasped with a minimum adaptation of the initial trajectory.
Previous work attempted to couple the uncertainty reduction and grasp execution in the framework of partially observable Markov decision processes (POMDPs). Although methods are advancing for continuous state (Porta et al. 2006; Bai et al. 2010; Brooks et al. 2006) and continuous action spaces (Porta et al. 2006; Murphy 2000), no POMDP planners yet scale to the high dimensional belief state and action spaces required for robot grasping. This is especially true for manipulators with great dexterity as they have a high number of degrees of freedom. Instead, actual robot implementations of the POMDP approach to active tactile grasping separate exploratory and grasping actions and plan how to sequence them (Hsiao et al. 2011b), typically by relying on a user-defined threshold to know when the belief is accurate enough to switch from gathering information to execution of a pre-computed grasp action. This approach fails to exploit the fact that, in tactile grasping, hand movements can both move towards the grasp configuration, and reduce uncertainty. They are most naturally performed concurrently, rather than sequentially. Furthermore, these approaches typically rely on constraining the belief space to Gaussian distributions. Extensions to non-parametric representations of belief (Nikandrova et al. 2013), typically result in intractable planning problems due to the high dimensionality of non-Gaussian parametrisation.
The work presented in the paper "Hypothesis-based Belief Planning for Dexterous Grasping" by Dr Claudio Zito and his colleagues at the University of Birmingham (UK) shows a novel formulation of dexterous manipulation that aims to exploit concurrency in exploratory and grasping actions for reach-to-grasp hand movements.
Previous approaches fail to exploit the fact that, in tactile grasping, hand movements can both move towards the grasp configuration, and reduce uncertainty. These two actions, reaching an object and reducing uncertainty, are most naturally performed concurrently, rather than sequentially.
The properties of the approach are that it:
1) tracks high-dimensional belief states defined over a 6D, non-Gaussian pose uncertainty;
2) efficiently plans in a fixed-dimensional space;
3) simultaneously gathers information while grasping, i.e., there is no need to switch between gathering information and grasping since the action space is the space of dexterous reach-to-grasp trajectories;
4) does not require a user-supplied mesh model of the object or a pre-computed grasp associated with the object;
5) copes also with non-convex objects, i.e., there are no limitations to the shape of the objects that it can successfully grasp.
This work builds our approach by combining the idea of hypothesis-based planning (HBP), initially proposed in (Platt et al. 2011), and our one-shot learning algorithm for dexterous grasping of novel objects (Kopicki et al. 2015). The hypothesis-based planner works as follows. Instead of planning directly in a high dimensional belief space, our plan is constructed on a fixed-dimensional, sampled representation of belief. In other words, the belief space is projected onto a set of competing hypothesis in the underlying state space. However, our implementation of the HBP algorithm extends the work in (Platt et al. 2011) in several directions. First, Platt’s formulation of the hypothesis-based planner is defined on a set of actions (i.e., movement constrained in the horizontal plane) that differs from the actual grasp (i.e. a pinch grasp with two paddles). In contrast, we formulate the problem on the same action space for each stage (i.e., dexterous reach-to-grasp trajectories). As a result,we do not require a user-supplied threshold over the current belief to estimate when to interrupt the information gathering phase and execute a pre-defined grasp. Another difference is that the observational model used in (Platt et al. 2011) relies on contactless sensors (i.e.laser sensors), while we maximise tactile observations for a dexterous robotic hand; and, finally, we do not make any assumptions about the model of the object to be grasped, in contrast to the original work that assumes a convex object (i.e. a box) aligned in front of the robot. On top of our hypothesis-based planner, our grasping algorithm enables us to learn a set of grasp contacts on a point-cloud object model (PCOM), directly obtainable from a depth camera, and to generalise to objects of novel shape. Therefore we do not require a mesh model of the object, and we can also generate grasps on target objects with incomplete point clouds. Hence our algorithm is exceptionally flexible in planning dexterous reach-to-grasp trajectories for novel objects.In order to link these two methods, hypothesis based planning and dexterous grasping of novel objects, we need to construct a representation of the belief space that will allow us to track pose uncertainty for a PCOM, in 6D, and cope with the non-Gaussian posterior. We do so by employing a non-parametric representation of the belief state defined as a kernel density estimator. Each kernel is a weighted pose of the target object inferred from visual data collected on the fly.
Although we plan for generating the most informative contact, by integrating the pose uncertainty in the planning phase, we obtained a more robust reaching trajectory that copes with the uncertainty since the early stages. As results, we double the proportion of times in which the robot reaches and grasps the object at the first attempt.
Our experimental results show that our planner, IR3ne, is more reliable than open-loop grasping from vision. We further show that IR3ne improves oversimple tactile re-planning in three ways: i) it doubles the proportion of times in which the robot reaches and grasps the object at the first attempt, ii) if an unexpected contact is generated along the trajectory, the contact provides more information about the object’s pose, and thus iii) it reduces the number of re-planning steps before converging to a grasp. Experiments in simulation and on a real robot confirm these attractive properties.
Porta, J.M., Vlassis, N., Spaan, M.T., Poupart, P. (2006). Point-based value iteration for continuous pomdps. Journal of Machine Learning Research 7 (Nov), 2329–2367
Bai, H., Hsu, D., Lee, W.S., Ngo, V.A. (2010). Monte-carlo value iteration for continuous-state pomdps. Algorithmic foundations of robotics IX, pp. 175–191. Springer
Brooks, A., Makarenko, A., Williams, S., Durrant-Whyte, H. (2006). Parametric pomdps for planning in continuous state spaces. Robotics and Autonomous Systems 54(11), 887–897
Murphy, K.P. (2000). A survey of pomdp solution techniques. Environment Science and Policy for Sustainable Development 2
Hsiao, K., Kaelbling, L.P., Lozano-P ́erez, T. (2011b). Robust grasping under object pose uncertainty. Autonomous Robots 31(2-3), 253–268
Nikandrova, E., Laaksonen, J., Kyrki, V. (2013). Towards informative sensor-based grasp planning. Robotics and Autonomous Systems pp. 340–354
Platt, R., Kaelbling, L., Lozano-Perez, T., Tedrake, R. (2011). Efficient planning in non-gaussian belief spaces and its application to robot grasping. Proc. of the Int. Symposium on Robotics Research
Kopicki, M., Detry, R., Adjigble, M., Stolkin, R., Leonardis, A., Wyatt, J. (2015). One shot learning and generation of dexterous grasps for novel objects. The International Journal of Robotics Research
#robotics #dexterous #grasping #human #movement #UoB #memnone #Claudio #Zito #Hypothesis #pose #uncertainty #efficient #planning #Gaussian #realrobots #boris #humanoid #manipulator #reasoning #AI #ML #science #robot