- Time: 12.00 - 13.00
- Location: Sir William Henry Bragg Building, GR.18 (register necessary to attend - see below)
- External URL: https://docs.google.com/forms/d/e/1FAIpQLSfoarUhkgCMZN2Lx-BIJj51edOD7KBSB6msvUlDiM3i-lyRmg/viewform
Planning a constrained obstacle-free motion for a robot can be computationally complex, often resulting in slow computation of inefficient motions. In this talk, we propose solutions to (1) speed up computation using Fog Robotics serverless computing, (2) compute time-optimized motions that incorporate pick-and-place gripper-aware constraints, and (3) warm start motion planning using deep learning. In the first part, Fog Robotics algorithms use serverless lambda computing to scale computational parallelism on demand. The proposed algorithms empirically scale to 100s of simultaneous lambda executions, allowing robots to compute motion plans that converge in a fraction of the time. In the second part, we address robot-automated warehouses, in which the ability to quickly compute and execute efficient pick-and-place motions is critical to addressing labor shortages, lowering costs, and increasing revenue. We introduce a grasp-optimized motion planner (GOMP) that incorporates time-optimization and constraints on degrees of freedom implied by the gripper design and grasp-analysis into a sequential quadratic program to reduce motion execution time. In the third part, to increase reliability and reduce wear on robot joints, we extend GOMP to limit and minimize jerk and train a deep neural network to rapidly generate trajectories that warm start the optimization. The deep-learning jerk-optimized GOMP (DJ-GOMP) computes trajectories in milliseconds, empirically speeding up computation by 300x. We demonstrate the effectiveness of both serverless and DJ-GOMP motion planners in simulation and on physical robots.
Jeffrey Ichnowski is a post-doctoral researcher in the RISE lab and AUTOLAB at the University of California at Berkeley. He researches algorithms and systems for high-speed motion, task, and grasp planning for robots, using cloud-based high-performance computing, optimization, and deep learning. Jeff has a Ph.D. in computational robotics from the University of North Carolina at Chapel Hill. Before returning to academia, he founded startups and was an engineering director and the principal architect at SuccessFactors, one of the world’s largest cloud-based software-as-a-service companies.