Organizers: Dieter Fox
Abstract: After providing a brief overview of the Sensor Systems group, I will present our recent work in robotics. I will introduce pretouch sensing, our term for in-hand sensing that is shorter range than vision but longer range than tactile sensing. I will review Electric Field Pretouch sensing, introduce Seashell Effect Pretouch, and discuss strategies for using pretouch sensing in the context of robotic manipulation. As an active sensing modality, pretouch requires a choice of "next view." Since the robot hand is used for both sensing and actuation, pretouch-enabled grasping also requires us to consider an exploration/execution tradeoff. Finally, I will outline several new robotics projects that are underway.
Abstract: Underwater robotics is undergoing a transformation. Recent advances in AI and machine learning are enabling a new generation of underwater robots to make intelligent decisions (where to sample ? how to navigate ?) by reasoning about their environment (what is the shipping and water forecast ?). At USC, we are engaged in a long-term effort to develop persistent, autonomous underwater explorer robots. In this talk, I will give an overview of some of our recent results focusing on two problems in adaptive sampling: underwater change detection and biological sampling. I will also present our recent work on hazard avoidance, allowing robots to operate in regions where there is ship traffic. Bio: Gaurav S. Sukhatme is a Professor of Computer Science (joint appointment in Electrical Engineering) at the University of Southern California (USC). He is currently serving as the Chairman of the Computer Science department. His recent research is in networked robots.
Biography: Dr. Sukhatme has served as PI on numerous federal grants. He is Fellow of the IEEE and a recipient of the NSF CAREER award and the Okawa foundation research award. He is one of the founders of the RSS conference and has served as program chair of all three leading robotics conferences (ICRA, IROS and RSS). He is the Editor-in-Chief of the Springer journal Autonomous Robots.
Abstract: Part 1: Linear equivalent of dynamic programming. We show that Bellman equation for dynamic programming can be replaced by just as simple linear equation for the so-called optimal ranking function, which encodes the optimal sequence via its greedy maximization. This optimal ranking function represents Gibbs distribution which minimizes the expected sequence cost given the entropy level (set by a temperature parameter). Each temperature level gives rise to a linearly computable optimal ranking function. Part 2: Predictive state representation with entropy level constraint. Building on part 1, we show that if one specifies the entropy level of the input's stochastic process, then its Bayesian inference for the purposes of optimal learning can be simplified greatly. We conceptualize an idealized nervous system that is an online input-output transformer of binary vectors representing the neurons' firing states, and we ask how one would adjust the input-output mapping optimally to minimize the expected cost. We will argue that predictive state representations could be employed by a nervous system. Part 3: Evidence of optimal predictive control of human eyes. We present evidence of optimal-like predictive control of human eyes in visual search for a small camouflaged target. To a striking degree, human searchers behave as if maintaining a map of beliefs (represented as probabilities) about the target location, updating their beliefs with visual data obtained on each fixation using the Bayes Rule, and moving eyes online in order to maximize the expected information gain. Some of these results were published in Nature.
Abstract: One clear direction for the near future of robotics makes use of the ability to build and keep up to date geometric models of the environment. In this talk I will present an overview of my work in monocular real-time dense surface SLAM (simultaneous localisation and mapping) which aims to provide such geometric models using only a single passive colour or depth camera and without further specific hardware or infrastructure requirements. In contrast to previous SLAM systems which utilised sparser point cloud scene representations, the systems I will present, which include KinectFusion and DTAM, simultaneously estimate a camera pose together with a full dense surface estimate of the scene. Such dense surface mapping results in physically predictive models that are more useful for geometry aware augmented reality and robotics applications. Crucially, representing the scene using surfaces enables elegant dense image tracking techniques to be used in estimating the camera pose, resulting in robustness to high speed agile camera motion. I'll provide a real-time demonstration of these techniques which are useful not only in robust camera tracking, but also in object tracking in general. Finally, I'll outline our latest work in moving beyond surface estimation to incorporating objects into the dense SLAM pipeline.
Abstract: Science-fiction robots can perform any task humans do and more. In reality, however, today's articulated robots are disappointingly limited in their motor skills. Current planning and control algorithms cannot provide the robot with the capacity for intelligent motor behavior - instead, control engineers must manually specify the motions of every task. This approach results in jerky motions (popularly stereotyped as “moving like a robot”) that cannot cope with unexpected changes. I study control methods that automate the job of the controls engineer. I give the robot only a cost function that encodes the task in high-level terms: move forward, remain upright, bring an object, etc. The robot uses a model of itself and its surroundings to optimize its behavior, finding a solution that minimizes the future cost. This optimization-based approach can be applied to different problems, and in every case the robot alone decides how to solve the task. Re-optimizing in real time allows the robot to deal with unexpected deviations from the plan, generating robust and creative behavior that adapts to modeling errors and dynamic environments. In this talk, I will present the theoretic and algorithmic aspects needed to control articulated robots using model-based optimization. I will discuss how machine learning can be used to create better controllers, and share some of my work on trajectory optimization. A preview of some of the work discussed in this talk can be seen here: https://dl.dropbox.com/u/57029/MedleyJan13.mp4 [a lower-quality version is also available on youtube: http://www.youtube.com/watch?v=t4JdSklL8w0]
Abstract: If we hope to build an intelligent agent, we have to solve (at least!) the following problem: by watching an incoming stream of sensor data, hypothesize an external world model which explains that data. For this purpose, an appealing model representation is a dynamical system. Sometimes we can use extensive domain knowledge to write down a dynamical system, however, for many domains, specifying a model by hand can be a time consuming process. This motivates an alternative approach: *learning* a dynamical system directly from sensor data. A popular assumption is that observations are generated from a hidden sequence of latent variables, but learning such a model directly from sensor data can be tricky. To discover the right latent state representation and model parameters, we must solve difficult temporal and structural credit assignment problems, often leading to a search space with a host of (bad) local optima. In this talk, I will present a very different approach. I will discuss how to model a dynamical system's belief space as a set of *predictions* of observable quantities. These so-called Predictive State Representations (PSRs) are very expressive and subsume popular latent variable models including Kalman filters and input-output hidden Markov models. One of the primary advantages of PSRs over latent variable formulations of dynamical systems is that model parameters can be estimated directly from moments of observed data using a recently discovered class of spectral learning algorithms. Unlike the popular EM algorithm, spectral learning algorithms are statistically consistent, computationally efficient, and easy to implement using established matrix-algebra techniques. The result is a powerful framework for learning dynamical system models directly from data.