Explore cutting-edge robotics research from the NVIDIA Robotics Lab in this Fall 2019 seminar presented by Arsalan Mousavian and Nathan Ratliff. Dive into topics such as robot manipulation, physics-based simulation, and robot perception as the speakers discuss NVIDIA's efforts to develop AI-powered robots for real-world applications in manufacturing, logistics, and healthcare. Learn about innovative techniques like grasp sampling, structured latent spaces, and operational space control. Discover how NVIDIA is working towards creating robots that can robustly manipulate the physical world and collaborate with humans. Gain insights into the challenges and advancements in areas such as grasp evaluation, refinement, and learning from human demonstrations.
Overview
Syllabus
Intro
MOTIVATION
OVERVIEW
GRASP SAMPLER
STRUCTURED LATENT SPACE
GRASP EVALUATOR
GRASP REFINMENT
TRAINING
ABLATION STUDIES
LEARNED SAMPLER VS GEOMETRIC SAMPLER Geometric Sampler
FAILURE CASES Some of the failures are because of hand-eye calibration error
EXTENSION TO CLUTTERED SCENES
EXAMPLE: CLEARING TABLE
REMOVING BLOCKER OBJECT
EXTENSION TO OBJECT PLACEMENT
CONCLUSIONS
OPERATIONAL SPACE CONTROL Start from a technique we know to be fundamentally reactive.
RIEMANNIAN MOTION POLICIES Encoding behavior into the fabric of the task space
TASK SPACE ARE RIEMANNIAN MANIFOLDS Generalize the notion of straight using Riemannian geometry
OBSTACLE DAMPING Damp velocity components in the direction of obstacles
LEARNING FROM HUMAN DEMONSTRATIONS
LEARNING RIEMANNIAN METRIC
Taught by
Paul G. Allen School
