IPAB Workshop - 07/03/2019

Speaker: Keyhan Babarahmati

Talk Title:  A Passivity-based Approach to State-dependent Variable Impedance Control

Talk abstract:  Robots have widespread use in manufacturing where they operate in highly structured environments with minimal human-robot interaction.   Recently, robots have been introduced to other industries (e.g.\ healthcare), requiring more dynamic tasks that cannot be fully characterized a priori. The impedance controller, as introduced by Neville Hogan, is a widespread technique enabling robots to interact with uncertain environments, within certain boundaries, by controlling the robot to act as an equivalent mechanical impedance or Mass-Spring-Damper. Nevertheless, stability of such systems highly depends on proper controller gains, which are difficult to tune for dynamic tasks (e.g.: polishing, locomotion, etc.) in unstructured environments that require adaptive trajectories and/or variable impedance gains. In particular, tasks where uncertain end-effector contact state may occur (e.g.: polishing, human robot collaboration, etc.) pose a challenge to the robots' controllers that rely on contact for ensuring system stability. Therefore, it is required to find control methodology that gives robots the capabilities of varying their controller gains w.r.t. aforementioned dynamic tasks whilst remaining passive, as passivity of the system ensures its stability both in free motion and when in contact.

Speaker: Floyd Chitalu

Talk title: Ostensibly-implicit BVH trees for parallel collision detection

Talk abstract:  I will talk about my recent work on parallel collision detection in physics-based simulations. In this work, we introduce a method for representing the bounding volume hierarchy (BVH) under the assumption of an implicit tree for fast and efficient collision detection. Using (binary) heap representation and generalized pointer-less traversal, we produce a post-processing free data structure with an implicit layout and predictable memory costs which scale linearly with the number of mesh primitives. Explicit node-connectivity and the need for padding memory to account for missing elements are alleviated by using a novel formulation for mapping the implicit index representation to compact memory locations. The result is that our data structure permits fast and efficient GPU based collision with low memory budget, storing only the necessary information to perform traversal operations. In particular, its mathematical simplicity allows for fast hierarchy construction achieving over 5× speedup compared to the state-of-the-art.