TThe workshop is in room 217-219.
|10:15-10:45||Tea Break / Poster Discussion|
|11:10-11:35||Michael Posa, Jiaji Zhou|
|12:30-13:30||Lunch Break / Poster Discussion|
|15:20-15:55||Tea Break / Poster Discussion|
|16:00-17:00||Panel Discussion & Open Forum|
- Todd Murphey, “Model-Based Control for Data-Driven Models of Contact”
Abstract: This talk will discuss the use of spectral methods to generate data-driven models of mechanical contact for control synthesis. Specifically, we use Koopman operators, an infinite-dimensional linear representation of a differential equation, to compute and update equations of motion based on experimental data. We then apply model-predictive control techniques to the resulting model, updating both the model and the resulting control in real-time as a function of data. The talk will include a brief overview of Koopman operators, simulated and experimental outcomes, and a discussion of opportunities and limitations that arise when using data-driven techniques to replace or augment analysis-based models.
- Jeannette Bohg, “Combining learned and analytical models for predicting the effect of contact interaction”
Abstract: One of the most basic skills a robot should possess is predicting the effect of physical interactions with objects in the environment. Traditionally, these dynamics have been described by physics-based analytical models which may be very hard to formulate for complex problems. More recently, we have seen learning-based approaches that can predict the effect of complex physical interactions from raw sensory input. However, it is an open question how far these models generalize beyond their training data.
In this talk, I propose a way to combine analytical and learned models to leverage the best of both worlds. The method assumes raw sensory data as input and the predicted effect as output. In our experiments, we compared the performance of the proposed model to a purely learned and a pure analytical model. Our results show that the combined method outperforms the purely learned version in terms of accuracy and generalization to interactions and objects not seen during training. Beyond these empirical result, I will also present an in-depth analysis of why the purely learned model has difficulties in capturing the dynamics of this task and how the analytical model helps. I will share the insights we gained on how to combine learned and analytical models.
- Ludovic Righetti, “Optimization and data-driven approaches for dynamic contact behaviors”
Abstract: Important progress has been made these past years for the planning (and
optimization) of multi-contact behaviors. However, current algorithms tend to suffer from high computational complexity and resulting plan executions are usually very sensitive to contact uncertainty. In this presentation, I will present our recent results in computing multi-contact motions using a convex relaxation of the centroidal momentum dynamics that allow us to come to near real-time computation times while keeping solutions very close to the original, non-convex problem. In a second part, I will show how data-driven methods can be used to better estimate contact modes and create controllers that perform well in uncertain and changing environments.
- Nathan Ratliff, “Exploiting optimization structure and optimality principles to embed dynamics and simplify contacts for motion optimization”
Abstract: Kinematic motion generation and optimization under non-linear dynamics are usually treated as distinct classes of problems with different computational complexity. Dynamics are tough, especially when contacts are involved, due to their non-linearity, the difficulty of computing derivatives, and their complimentarity constraints. But Newton’s second law (f = ma) suggests that there’s a close connection between kinematics and dynamics. Likewise, Lagrange multipliers appear in kinematic constraints analogous to forces, in the same way they do as real forces in the constrained optimality principles of classical mechanics that the dynamic equations of motion are derived from. Are they really that different? Complimentarity is frequently cited as a primary source of complexity in optimizing contact dynamics. And it’s true, complimentarity makes optimization really hard, even combinatorically NP-hard. But this has been a problem for decades: every inequality constrained optimization problem is burdened with complimentarity, it’s one of the KKT conditions. Complimentarity is what makes inequality constrained optimization so much more difficult than unconstrained optimization. And because of that, optimization researchers have focused their efforts toward developing very powerful and efficient second-order inequality constrained optimization solvers designed explicitly to solve this problem of complimentarity. Interior point methods, for instance, are a series of relaxations to the KKT conditions’ complimentarity constraint prescribed to circumvent that condition’s computational complexity. There’s reason to believe that non-linear dynamics and contacts could be handled more efficiently in motion optimization. In this talk, I’ll build on these observations. In the first half, I’ll step first through a series of results exploring embedding dynamical quantities into kinematic motion optimization and exploiting Lagrange multipliers at surface contacts. And in the second half, I’ll discuss some preliminary work on developing a computational theory of motion optimization with embedded dynamics designed to better exploit optimality principles and second-order solvers for constrained optimization. The aim of this new line of research is to develop a new framework for motion optimization with contacts to close the computational gap between kinematic motion generation and optimization with dynamics and contacts.
- Kris Hauser, “Toward data-driven contact mechanics”
Abstract: Mathematical tools from contact mechanics are used throughout manipulation planning, control, legged locomotion, and physics simulation, but are largely based on idealizations that do not adequately capture many real world phenomena. Alternatively, measurements or observational data could be used (i.e., via machine learning) to make high-fidelity predictions of behaviors that are hard to capture in idealized models, such as intimate contact, deformation, and adhesion. It remains challenging to generalize empirical data across vastly different scenarios. This talk argues that data-driven contact mechanics — using mathematical modeling to compliment data — can lead to predictions that are both accurate and general.
- Alberto Rodriguez, “Capabilities that I wish my robots would have”
Abstract: In this presentation, I will start by describing a list of capabilities that I wish my robots would have. These are not problems that I know how to solve, but rather keep me busy, and that I believe are essential to any robotic manipulation system which hopes to be called skilled and practical. I will share some ideas and ask for yours. In the second part of the presentation I will describe current efforts
in my group to enable model-based planning of frictional in-hand manipulation by exploiting contacts with the environment, and on real-time state estimation fusing tactile and vision sensing, leveraging techniques for dynamic smoothing developed by the SLAM community.
- Aaron Johnson, “Hybrid Systems Modeling for Robotic Systems with Contact” Abstract: In this talk, I will discuss several modeling considerations for analyzing robotic systems involving changing contact conditions. The models considered are, like all models, only approximations but fit well to the domain of most manipulation and self-manipulation robots with a limited number of contacts that persist over time. In particular I will use a number of common physical assumptions such as plastic impact or massless legs and show some of the technical aspects of a hybrid dynamical system that includes these assumptions and also satisfies a number of consistency properties.
- Bernardo Aceituno-Cabezas, Hongkai Dai, Carlos Mastalli, Michele Focchi, Andreea Radulescu, Darwin G. Caldwell , José Cappelletto , Juan C. Grieco , Gerardo Fernández-López, and Claudio Semini. A Mixed-Integer Convex Formulation for Simultaneous Contact, Gait and Motion Optimization on Multi-Legged Robots. [abstract]
- Jonathan Arreguit , Salman Faraji , and Auke J. Ijspeert. Multi-contact motion planning with a five-mass template model and vector-based equations. [abstract]
- Silvia Cruciani and Christian Smith. Pivoting Objects with a Parallel Gripper Using Controlled Slip and Inertial Forces. [abstract]
- Wesley Guo and Kris Hauser. Fast Equilibrium Testing with Empirical End Effector Wrench Spaces and Compliance in a Climbing Robot. [abstract]
- Sabrina Hoppe, Zhongyu Lou, Daniel Hennes, and Marc Toussaint. Deep Learning for Manipulation with Visual and Haptic Feedback. [abstract]
- Sheng Li, Tianxiang Zhang, Guoping Wang, Hanqiu Sun, and Dinesh Manocha. Multi-contact Frictional Rigid Dynamics using Impulse Decomposition. [abstract]
- Zachary Manchester and Scott Kuindersma. Variational Contact-Implicit Trajectory Optimization. [abstract]
- Romeo Orsolino, Michele Focchi, Carlos Mastalli, Hongkai Dai, Darwin G. Caldwell and Claudio Semini. The Actuation-consistent Wrench Polytope (AWP) and the Feasible Wrench Polytope (FWP). [abstract]
- Jian Shi and Kevin M. Lynch. In-hand Sliding Manipulation with Spring-Sliding Compliance. [abstract]
- Haoran Song, Michael Yu Wang, and Kaiyu Hang. Grasping on Contact Primitives by Fingertip Surface Optimization. [abstract]
- Markku Suomalainen and Ville Kyrki. Towards environment-aware compliant motion primitives for assembly. [abstract]