Events

CS Colloquium: Integrative Learning, Modeling, Perception, and Planning Solutions for Modern Robotics

Speaker: Kostas Bekris (Rutgers University)

Location: 60 Fifth Avenue, Room 150

Date: Wednesday, October 30, 2024

Robots are increasingly deployed in the real-world but there are still critical gaps that limit their adaptability, robustness and safety. This talk will discuss how tighter integration of perception, learning and planning for vision-driven robot manipulation can address gaps that lie at the interface of these subdomains, which must be viewed holistically in robotics. We will demonstrate how progress in 3D robot perception, appropriate representations, tighter closed-loop operation and compliant end-effectors allow full-stack robot manipulation systems in applications, such as robot packing, assembly under tight tolerances and constrained placement. The talk will also focus on the Sim2Real gap in the context of novel, soft-rigid robotic platforms, such as tensegrity robots for uneven terrain locomotion. Modeling and controlling is complicated for such robots due to high-dim., complex dynamics and the sim2real gap. Reinforcement learning (RL) is promising but limited due to high data needs. Training RL in simulation is also blocked by the sim2real gap. We will present differentiable engines for tensegrity robots, which can be trained with few real trajectories, to provide accurate-enough models to define controllers transferable back to the real system. As learning plays an increasing role in robot control, it also becomes critical to identify the conditions under which learned controllers can successfully control robots to a desired state. We have been exploring topological tools for computing Region of Attraction (RoA) of black-box robot controllers. These tools achieve reduced data requirements relative to alternatives in approximating RoAs, and combined with machine learning methods, can be applied for RoA identification of high-dim. robotic systems.