Keynotes and Tutorials

Keynotes

Keynote 1:  Leslie Pack Kaelbling    [Google Scholar]

Title: Doing for our robots what evolution did for us

Abstract:

Within robotics and AI, we have made great strides in planning, learning, and reasoning.  But these improvements are moving forward largely independently, resulting in competing strategies for solving similar classes of problems.  A better strategy would be to figure out which aspects of our techniques for planning and reasoning can leverage learning to be more effective, robust, and sample-efficient.  I will discuss many different roles that learning can play in an intelligent robotic system and ways in which they can be scaffolded by architectural and algorithmic commitments.

Keynote 2: Danica Kragic     [Google Scholar]

Title: Robot systems that act, interact and collaborate

Abstract:

The integral ability of any robot is to act in the environment, interact and collaborate with people and other robots. Interaction between two agents builds on the ability to engage in mutual prediction and signaling. Thus, human-robot interaction requires a system that can interpret and make use of human signaling strategies in a social context. In such scenarios, there is a need for an interplay between processes such as attention, segmentation, object detection, recognition and categorization in order to interact with the environment. In addition, the parameterization of these is inevitably guided by the task or the goal a robot is supposed to achieve. In this talk, I will present the current state of the art in the area of robot perception and interaction and discuss open problems in the area. I will also show how visual input can be integrated with proprioception, tactile and force-torque feedback in order to plan, guide and assess robot's action and interaction with the environment. For interaction, we employ deep generative models that makes inferences over future human motion trajectories given the intention of the human and the history as well as the task setting of the interaction. With help predictions drawn from the model, we can determine the most likely future motion trajectory and make inferences over intentions and objects of interest.

Keynote 3: Joelle Pineau     [Google Scholar]

Title: Reproducibility, Reusability, and Robustness in Deep Reinforcement Learning

Abstract:

In recent years, significant progress has been made in solving challenging problems across various domains using deep reinforcement  learning. However reproducing results for state-of-the-art deep RL methods is seldom straightforward. High variance of some methods can make learning particularly difficult when environments or rewards are strongly stochastic. Furthermore, results can be brittle to even minor perturbations in the domain or experimental procedure. In this talk, I will discuss challenges that arise in experimental techniques and reporting procedures in deep RL, and will suggest methods and guidelines to make future results more reproducible, reusable and robust.

Keynote 4:  Daisuke Okanohara     [Google Scholar]

Title: Robots for solving real-world tasks

Abstract:

I will introduce two example tasks among our activities to show the possibilities and difficulties in using robots for solving real-world

tasks. The first is a picking task. Recent studies and challenges (e.g. Amazon picking challenges) have shown the remarkable progress on using robots for picking tasks. In particular deep learning-based recognition and planning allow for the construction of robust and

scalable picking systems. To deploy these systems in the industry, robots need to pick over several million distinct items even including

unknown items. To solve the picking in this scale we require robust, generalizable, easy-to-prepare solutions. The second is using unconstrained spoken language for instruction. Using spoken language for interaction has two merits. First, a user is not required to learn a special skill (e.g. programming) to perform a task. Second, spoken language can contain various information such as object description, orientation, space and time. We show that a spoken language interface is easy to use, and can handle many different situations.

Keynote 5: Marc Toussaint     [Google Scholar]

Title: Symbols for Manipulation and Physical Reasoning

Abstract:

I will start with discussing general purpose physical reasoning and manipulation, the idea of differentiable simulations, and limitations of this idea. My focus is then on planning and optimization approaches to physical reasoning and sequential manipulation planning. In both cases, abstractions or symbols are essential to formulate bounds and heuristics, which are at the core of efficient solvers. In my view, this research area sheds new light on the role of abstractions for reasoning, and eventually for learning. While the focus of this talk is on planning and reasoning, I will argue, and give examples, that an understanding of these problems is essential also for (sample-efficient) learning.

Keynote 6: Masahiro Fujita     [ResearchGate]

Title: Entertainment Robot Aibo (Tentative)

Abstract:

Sony made the announcement of a new entertainment robot, aibo, in November 11, 2017.

From the original version of AIBO sold in 1999, it is almost 20 years past. I will describe the new aibo comparing with the original aibo. In addition I will introduce Sony's robotics activities until 2018.

Tutorials

Tutorial 1: Jean-Jacques Slotine      [Google Scholar]

Title : Combining physics and neural networks for stable adaptive control and system identification

Abstract:

Concurrent learning and control in dynamical systems is the subject of adaptive nonlinear control. We discuss systematic algorithms in this context, both vintage and recent. 

The algorithms, applicable to both adaptive control and system identification, have guaranteed stability and convergence properties based on Lyapunov theory and contraction analysis. They easily combine information from physics (e.g., in robotics) with network-based representations, can accommodate local minima or redundancy, and can be made explicitly robust to unmodeled uncertainties. A key enabling aspect of this performance is that adaptation and learning occur on a "need-to-know" basis, in the sense that the system learns just enough to achieve the desired real-time control or identification task.

The algorithms can be extended to exploit dynamic representations based on multilayer or deep networks, with similar stability and

convergence guarantees. They can also exploit both physics-based and local basis functions for real-time computational efficiency.

Tutorial 2: Benjamin Recht      [Google Scholar]

Title : Optimization Perspectives on Learning to Control

Abstract:

Given the dramatic successes in machine learning over the past half decade, there has been a resurgence of interest in applying learning techniques to continuous control problems in robotics, self-driving cars, and unmanned aerial vehicles. Though such applications appear to be straightforward generalizations of reinforcement learning, it remains unclear which machine learning tools are best equipped to handle decision making, planning, and actuation in highly uncertain dynamic environments.

This tutorial will survey the foundations required to build machine learning systems that reliably act upon the physical world. The primary technical focus will be on numerical optimization tools at the interface of statistical learning and dynamical systems.  We will investigate how to learn models of dynamical systems, how to use data to achieve objectives in a timely fashion, how to balance model specification and system controllability, and how to safely acquire new information to improve performance. We will close by listing several exciting open problems that must be solved before we can build robust, reliable learning systems that interact with an uncertain environment.

Tutorial 3 : Emo Todorov      [Google Scholar]

Title : Sensorimotor intelligence via model-based optimization

Abstract:

Model-free reinforcement learning has produced surprisingly good results for a brute-force method. However it appears to be reaching an asymptote that is not competitive with model-based optimization. Furthermore it is mostly limited to simulation where a model is available by definition. So we might as well take full advantage of that model, and reserve model-free methods for fine-tuning on real data.

In this tutorial I will discuss state-of-the-art methods that become available once we (admit that) have a model. As with any other form of optimization, the single most important ingredient is having access to analytical derivatives. This is standard in supervised learning for example, but general-purpose physics simulators are difficult to differentiate. Nevertheless this is now possible in MuJoCo as well as some more limited simulators, opening up possibilities for much more efficient optimization. Another essential ingredient in the control context is inverse dynamics. This enables trajectory optimization methods where the consistency between states and controls no longer needs to be enforced numerically, and instead one has to enforce under-actuation constraints which are lower-dimensional. Another challenge, specific to problems with contact dynamics, is that contacts result in very complex optimization landscapes that can be difficult to navigate even for a full Newton method. Unlike the situation in neural networks where saddle points appear to be the problem, here the problem is harder: the gradient is large yet it changes rapidly in non-linear ways that are not captured by the Hessian (and we don't have 3rd-order methods). This can be alleviated using continuation methods, where the physics model is smoothed early in optimization and gradually made harder while tracking the solution. The algorithm best suited for this problem class is Gauss-Newton. Putting all the ingredients together, one iteration of trajectory optimization can be performed in a few milliseconds on a single computer. The same machinery can be used to solve state estimation problems and system identification problems, in addition to control problems. This is done by modifying the cost function and keeping everything else the same.

A down-side of this framework is that it involves more mathematics, physics, optimization and software engineering than what the community has gotten used to. A possible solution is to produce software that does it automatically, leaving parameter tuning to the user. We are in the process of developing such software (called Optico) and will show demos at the tutorial.

Tutorial 4:  Masashi Sugiyama       [Google Scholar]

Title: Machine learning from weak supervision---Towards accurate classification with low labeling costs

Abstract:

Recent advances in machine learning with big labeled data allow us to achieve human-level performance in various tasks such as speech recognition, image understanding, and natural language translation. On the other hand, there are still many application domains---including robotics---where human labor is involved in the data acquisition process and thus the use of massive labeled data is prohibited.  In this tutorial, I will introduce recent advances in classification techniques from weak supervision, including classification from two sets of unlabeled data, classification from positive and unlabeled data, and a novel approach to semi-supervised classification.