Interactive Perception for Learning the Dynamics of Articulated Objects

Manipulating articulated objects is an important skill for robots operating in human environments. We propose to learn a physical model of the dynamics of articulated objects to accurately predict the motion of the object. Being aware of the dynamic effects of its actions, the robot no longer needs to maintain a firm grasp of the handle over the full course of the manipulation, which allows for one-pointcontact manipulation or early release. This ability reduces the degrees of freedom required of the manipulator and allows for high speed execution. We present an approach to learn the objects’ dynamics from sensor observations of the moving door. The observations can incorporate information from force sensing or depth measurements, as obtained from a laser range scanner. Our method allows the robot to interactively learn the door dynamics, updating the learned model from observations gathered during manipulation. We devise an algorithm to predict the dynamic behavior of doors within the first manipulation, which allows the robot to bootstrap the model itself. Current approaches to robotic manipulation of articulated objects do not make use of explicit knowledge about the dynamics of the object. Most approaches assume the manipulation execution to be quasi-static, i.e. slow enough that the inertial forces are negligible. This substantially reduces the execution speed for systems that do not maintain a closure grasp of the handle over the course of the manipulation [1]. Approaches in which the robot maintains such a grasp are robust to small inertial forces but require specialized controllers [2] or high dimensional motion planning [3] to avoid lateral forces between manipulator and handle. In all cases, the object needs to be released at rest, which means the end effector needs to be in contact until the desired position is reached. In general, this is a challenging problem, particularly for robots with limited reachability or low number of joints. We propose to take advantage of the dynamics of articulated objects. We learn the model of the object’s dynamics, i.e. mass or moment of inertia, and the deceleration of the object by friction and air drag with respect to velocity and position. The observations can be acquired during manipulation using force sensing capabilities of the manipulator or using a laser range scanner or depth camera. For a door opening task, we extract the state θ from the mean p and principle component axes m,n of the depth measurements in a small bounding polygon placed at the initial guess for the door location. The angle is computed from the door normal. We adaptively grow the boundary by tracking the door (see Figure 1b). The hinge location h is estimated from the sequence [pt,nt] using a least squares approximation. We use locally weighted regression to robustly fit a second order polynomial to the time series [θt]. The velocity ωt and deceleration αt can be obtained by the first and second derivative of the local fit. We use a This work has partly been supported by the European Commission under the contract number FP7-ICT-248258-First-MM, the National Science Foundation under grant CCF-1208468, FRIAS, University of Freiburg and the DARPA under W15P7T-12-1-0002 (a) Being able to predict the dynamic behavior of the door, the robot can swing it open precisely to a desired state, here 60◦. The manipulation time from touch to release of the door is approximately one second. Initial estimate of the door location Door state