Speakers
Kenji Tahara
Title: Object Grasping and Manipulation under Inadequate Sensory Information
Abstract: We propose a control method for a multifingered robotic hand that can adjust the appropriate graspingforce against the external force applied to the object. We also propose a control method for a multifingeredrobotic hand that is robust for time-latency of sensory information and can correctly control the orientation of theobject using a soft visual tactile sensor.
Maximo Roa
Title: Environment-Aware Manipulation Planning for a Variable Stiffness Hand
Abstract: This talk deals with the problem of planning grasp strategies for constrained and cluttered scenarios, such as pick-from-the-bin cases. To successfully exploit the environmental constraints, we use the CLASH hand, a compliant hand that can vary its passive stiffness. The overall planning strategy considers multiple factors, including the arm capabilities, the possible environmental constraints that the hand can use for grasping even with uncertainties, the required stiffness setting for the hand, and the confidence on the learning-based vision input for detecting the objects.
Zoe Doulgeri
Title: Total Singulation of a target object in clutter with non-prehensile manipulation primitives: a modular RL approach
Abstract: A modular RL with continuous actions that allows to combine different policies which may be produced independently by RL, supervised learning or algorithmic design is shown to achieve effectively singulation of a target object from its surrounding clutter in different environments. Prior knowledge is effectively incorporated into learning by shaping the action space and properly selecting the state representation so that it focuses only on the singulation task instead of learning other tasks like obstacle avoidance. Non-prehensile actions include exertion of forces on the object sides or object top surfaces. The former relies on kinematic controls while the latter is achieved by devising control schemes that take into account analytic contact physics so that it can be robustly transferred in a real world environment.
You Zhou
Title: Learning compliance adaptation for contact-rich manipulation
Abstract: Compliant robot behavior is crucial for contact-rich manipulation tasks. The talk will present a novelapproach for learning predictive models of force profiles required for single handed and bimanual contact-richmanipulation tasks based on Bidirectional Gated Recurrent Units (Bi-GRU) and an adaptive force/impedancecontroller.
Oliver Kroemer
Title: Modularity in Learning to Grasp
Abstract: When addressing a challenge as large as robot grasping, it is important to break the problem down into smaller parts and as researchers discuss how to do this division effectively. During my talk, I will be presenting recent work on breaking down the requirements of grasps into modular parts, as well as employing modularity in control for grasping and manipulating objects.
Oliver Brock
Title: Towards simple and robust manipulation: Let the physics deal with physics
Abstract: Contact dynamics, high-dimensional configuration spaces, uncertain execution, poor sensing: dexterous manipulation is a difficult problem. It becomes much simpler if the challenges are distributed appropriate across morphological computation and explicit computation.
Franziska Meier
Title: Differentiable and Learnable Robot Models
Abstract: In the past, researchers have strictly distinguished analytical/physical models from data-driven learned models. Recently, a lot of work has explored the line between physical and data-driven models. In this talk, I will discuss our recent progress on building differentiable and learnable robot models which can merge varying amounts of structure with data-driven approaches.
Ken Goldberg
Title: Exploratory Grasping
Abstract: Inspired by infants that repeatedly attempt to grasp a toy until they can learn reliable ways to grasp it, we consider a novel problem: Exploratory Grasping, where a robot is presented with an unknown object and learns to reliably grasp it by repeatedly attempting grasps and allowing the object pose to evolve based on grasp outcomes. The objective is for the robot to explore grasps across different object poses to reliably grasp the object from any of its stable resting poses. We formalize Exploratory Grasping as an MDP in which the robot attempts grasps in the current stable pose, and if a grasp is successful, the robot lifts and releases the object to sample a new random stable pose. If a grasp is unsuccessful, the object either remains in the same stable pose or topples into a new stable pose as a result of the perturbation from the failed grasp.
Karl Van Wyk
Title: Geometric Fabrics for Efficient Behavior Encoding
Abstract: Generalization arises from strong compression of behavior. For instance, cost functions represent compact encodings, and when they can be decoded efficiently (eg fast A* planning) we observe strong generalization when behavior is encoded as costs (eg inverse optimal control). However, for many higher-dimensional problems encountered in collaborative robotics, these cost based encodings are challenging to decode efficiently (planning and global motion optimization are difficult) making cost-based representations often ineffective or impractical. This talk describes a new behavior encoding medium called geometric fabrics which builds on the observation that behavior can be encoded directly into the fabric of space by, for instance, geometrically stretching it. This encoding medium is both a compact encoding and efficient to decode in practice leading to strongly generalizing real-world systems.