Adrian Zwiener

Mobile Manipulation is versatile field in robotics involving vision, planning and control. A typical task in mobile manipulation is grasping, which involves three phases: perception, motion planning, and trajectory tracking. Usually, motion planning and trajectory tracking employ visual 3D obstacle avoidance and self-collisions are also avoided. Mobile manipulation in human environments requires high levels of safety, dexterity and robustness since failure rates and accidents have to be minimized. For safe and robust systems it is not sufficient to rely on obstacle avoidance alone because obstacle localization can be imprecise and can errors accumulate with small deviations in the tracked trajectory. Besides, during grasping, distances to objects are very narrow. Research has shown that with joint force/torque sensing the safety of manipulators and grasping success can be enhanced. Consequently, our research goal is to incorporate tactile feedback through joint torque sensing in the grasping process of mobile manipulators.

For the purpose of this research, our chair is equipped with three manipulators. For stationary manipulation we have one Franka Emika Panda arm. To research mobile manipulation we use two Jaco2 manipulators produced by Kinova. Both Jaco2s can be fixed to a Metralabs Scitos G5 platform or one each to two Scitos G5 or the smaller omnidirectional robots.

Manipulation by mobile service robots in human environments requires a high level of safety, therefore a robot must be able to detect external contacts, in order to interact with objects and humans reliably. We present a model-based Machine Learning (ML) approach to detect and localize external contacts on a 6 DoF serial manipulator. This approach only requires the use of proprioceptive sensors (joint positions, velocities and 1Dl joint torques), which are standard in modern robot arms. Good results are obtained with Random Forests (RFs) and Multi-Layer-Perceptrons (MLPs) leading to a precise localization of the contact link and its orientation. Apart from the link in contact and the orientation of the force, RFs and MLPs are also able to differentiate between contact points on the same link and orientation but with different distances to the joint axis. We experimentally verify this approach on simulated and real data obtained from the Kinova Jaco 2 manipulator and compare it to an optimization based approach.

Fig.3 Labelled confusion matrix. The label 0 indicates no contact. Points with lower numbers are closer to the base. Left: simulation results: DIRECT optimization with max. 90 iterations and relative param. tolerance of 1.6%. Centre: Simulation results: MLP 1 hidden layer 100 neurons and torque features. Right: Real data results of a RF, max. depth = 20, using acceleration and torque features.

Demonstration of the contact point loacalization on the Jaco2:

De Luca et al. introduced a generalized momentum observer to estimate the torques due to an external source. This observer can be used to implement a collision reflex control law. Originally, this reflex was implemented using torque control. However, a similar control law can be used with velocity control which is currently used for the Jaco2.

Fig.4 Experiment: Jaco 2 moves against an object and moves back to the start configuration. The red shaded area indicates a contact. The green shaded and dashed framed areas indicate movement.

Demonstration of the collision reflex.