Max Planck Symposium on Autonomous Systems: State-of-the-Art and
Organized by the Chemistry, Physics & Technology Section (CPTS) of the Max Planck Society
January 30-31, 2009, Tübingen, Germany
This international symposium brings together leading researchers in the broad area of autonomous systems, to share their perspective and vision about the future of the field.
January 29, 2009
07:00 PM: "Get together" at Hotel Krone, Big Club Room (Speakers and Committee members only)
January 30, 2009
09:00 AM: Opening by Werner Hofmann, Chair of the CPT Section
09:15 AM: Reinforcement Learning, Control, Game Theory
The Simultaneous Localisation and Mapping (SLAM) problem as originally posed is now well understood; one might say "solved" in small local areas. There is however much to be done in producing useful maps and representations of truly large workspaces and in obtaining sustained and robust operation within them. This talk will describe work on these issues and will review work on appearance based navigation and mapping over large scales and in particular how it solves the "loop closing" problem which plagues online infrastructure-free navigation algorithms. We shall also review work on detailed acquisition and semantic labelling of workspaces using both laser and appearance information. The "L" is done the "M" needs work.
Value-function based reinforcement learning (RL) has achieved remarkable successes during the last two decades. The strengths of this approach comes from combining sampling, dynamic programming and function approximation in a single unified framework. A basic step of most RL algorithms is to estimate the value function underlying some target policy. However up until recently this approach was believed to be unsound when the policy generating the data is different from the target policy. In this talk I will survey recent promising results that show how this barrier can be sidestepped.
Machine Learning is a core enabling technology for building autonomous systems that adapt to changing environments and react to unexpected situations. Unlike the statistical approach to learning, which is mostly based on the unrealistic assumption of stationary data sources, game theory provides a radically different model of the interaction between a learning agent and the environment. In this model, adaptation to arbitrary and even adversarial sources is made possible, and rigorous performance guarantee can be provided under no statistical assumptions on the environment. In the talk we will delineate the theoretical foundations of game-theoretic learning, give some examples of concrete algorithms, and suggest some interesting research directions.
This talk will describe the STAIR home assistant robot project, and the
satellite projects that led to key STAIR components such as (i) robotic
grasping of previously unknown objects, (ii) depth perception from a
single still image, and (iii) multi-modal robotic perception.
Since its birth in 1956, the AI dream has been to build systems that exhibit broad-spectrum competence and intelligence. STAIR revisits this dream, and seeks to integrate onto a single robot platform tools drawn from all areas of AI including learning, vision, navigation, manipulation, planning, and speech/NLP. This is in distinct contrast to, and also represents an attempt to reverse, the 30 year old trend of working on fragmented AI sub-fields. STAIR's goal is a useful home assistant robot, and over the long term, we envision a single robot that can perform tasks such as tidying up a room, using a dishwasher, fetching and delivering items, and preparing meals.
In this talk, I'll describe our progress on having the STAIR robot fetch items from around the office, and on having STAIR take inventory of office items. Specifically, I'll describe: (i) learning to grasp previously unseen objects (including unloading items from a dishwasher); (ii) probabilistic multi-resolution maps, which enable the robot to open/use doors; (iii) a robotic foveal+peripheral vision system for object recognition and tracking. I'll also outline some of the main technical ideas---such as learning 3-d reconstructions from a single still image, and reinforcement learning algorithms for robotic control---that played key roles in enabling these STAIR components.
One of the grand challenges in artificial intelligence is to create autonomous humanoid robots. The anthropomorphic shape of a humanoid should enable operation in environments designed for humans, the utilization of human tools and interfaces, and the natural use of human gestures and non-verbal communication. Fundamentally, an intelligent humanoid should be a truly "general-purpose" robot, able to accomplish any task a real human can. This talk will discuss the challenge of motion autonomy for humanoid robots and present an overview of several autonomous motion planning methods designed for application tasks involving navigation, object grasping and manipulation, footstep placement, and full-body motions. Experimental results on several humanoid platforms around the world will be shown, along with some new efforts in "mobile manipulation". Finally, the long-term prospects for the future development of robot autonomy and search-based AI will be discussed.
Over the last decade, motion planning algorithms have been developed to solve complex geometric problems and have contributed to advances in industrial automation and autonomous exploration, but also in diverse fields such as graphics animation and computational structural biology. This talk begins by reviewing the state-of-the art in sampling-based motion planning with emphasis on work for systems with increased physical realism. Then recent advances in planning for hybrid systems will be described, as well as the challenges of combining formal methods and planning for creating safe and reliable systems.
There is increasing interest, activity, and urgency to build robots capable of performing manipulation tasks with human-like competency in open and uncontrolled environments. I will present work toward this objective in the areas of motion planning, perception, and manipulation. This workfocuses on challenges that arise when scope and complexity of tasks cannot be anticipated by a programmer a priori but instead have to be assessed and addressed by the robot during task execution in an autonomous fashion. Overcoming these challenges will enable important novel applications of autonomous systems, ranging from healthcare to household robotics, manufacturing, or planetary ex ploration.
Mobile robots and mobile manipulation systems are becoming widely deployed, thanks to their increasing utility and affordability. They have started to carve out niches where there is a value proposition for users in the military, the home, and industry. This presentation will describe some of the current iRobot systems that have benefited from these trends. With thousands of ground robots deployed for tasks like IED neutralization, where they save lives every day, the question from the users is no longer whether there is a need for robots, but when more capabilities will be available. To deliver these capabilities, increased levels of autonomy are essential. This presentation will review our research efforts to enable increased levels of autonomy for much needed advances in force multiplication, multi-robot control, HRI, and manipulation.
In this talk I will discuss the challenges of building brains and bodies to create mobile autonomous systems that can interact in new ways with the physical world, on the ground, in water, and in the air. I will focus on recent progress in Autonomous Mobile Networks, which are distributed ad-hoc networks of robots that can sense, actuate, compute and communicate with each other using point-to-point multi-hop communication. The nodes in such networks include static sensors, mobile sensors, robots, animals, and humans. Such systems combine the most advanced concepts in perception, communication and control to create computational systems capable of large-scale interaction with the environmnet, extending the individual capabilities of each network component to encompass a much wider area, range of data, and control capabilities.
The perception-action cycle was described by the neuroscientist JM Fuster as the circular flow of information that takes place between the organism and its environment in the course of a sensory-guided sequence of behavior towards a goal. When the environment is (asymptotically mean) stationary, the efficiency of the cycle is determined by the ability of the organism to efficiently extract information from the past that is valuable for the organism in the future, on multiple time scales. This observation suggests an intriguing rigorous analogy between the perception action cycle and Shannon's classical model of communication. I will present this analogy and discuss some of its consequences for optimal biological adaptation and performance. More specifically, I will present some recent application of this theoretical framework to auditory perception.
In this talk, we tackle a fundamental problem that arises when using sensors to monitor the ecological condition of rivers and lakes, the network of pipes that bring water to our taps, or the activities of an elderly individual when sitting on a chair: Where should we place the sensors in order to make effective and robust predictions? Such sensing problems are typically NP-hard, and in the past, heuristics without theoretical guarantees about the solution quality have often been used. In this talk, I will present algorithms which efficiently find provably near-optimal solutions to large, complex sensing problems. Our algorithms are based on the key insight, that many important sensing problems exhibit submodularity, an intuitive diminishing returns property: Adding a sensor helps more if we have placed few sensors so far, and less if we have already placed many sensors. In addition to identifying most informative locations for placing sensors, our algorithms can handle settings, where sensor nodes need to be able to reliably communicate over lossy links, where mobile robots are used for collecting data or where solutions need to be robust against adversaries and sensor failures. I will present results applying our algorithms to several real-world sensing tasks, including environmental monitoring using robotic sensors, activity recognition using a built sensing chair, and a sensor placement competition. I will conclude with drawing an interesting connection between sensor placement for water monitoring, and the problem of selecting blogs to read in order to learn about the biggest stories discussed on the web. This talk is primarily based on joint work with Andreas Krause.
I will present some recent work on computational algorithms in robotics, and discuss whether they can perhaps provide insight into how the brain may accomplish similar tasks. In particular, I will discuss specific algorithms used for vision, motor control, localization, and navigation in robots. An overarching theme that emerges from these examples is the need to properly account for uncertainty and noise in the environment. I will show how current machine systems approach this problem, and hope to spur discussion about related processes among neurons and in the brain.
Recent work in machine learning has shown that nonparametric sufficient statistics can be computed efficiently and lead to powerful algorithms for visualization, matching, clustering, independent component analysis, feature selection, and can be seen as a unifying principle for a number of inference algorithms ranging from Gaussian Processes to Conditional Random Fields. In this work I outline how these techniques will help with addressing fundamental challenges in inference: the move to multicore and distributed computing and the advent of collaborative and customized estimation applications as they are present in spam filtering, recommender systems, distributed caching, network security, and personalized web search. The latter are all manifestations of autonomous systems with machine learning being a key ingredient in their successful deployment.
Mapping, navigation and other perception-based functions are essential to intelligent operation of autonomous vehicles. In complex outdoor environments; on unstructured terrain, in forests, underwater or in air, for example; the problem of perception is compounded by the infinite variety of possible features and objects, by generally large scales and by potentially poor environmental conditions. This talk will look at a number of theoretical and practical approaches to addressing this challenge. At the sensor level we consider and demonstrate how multiple sensor modalities can be used to provide richer features and maps which are more robust to environment conditions. At the algorithm level, we consider various machine learning and appearance-based approaches to mapping which can deal with richer and denser feature sets. At the functional mapping and navigation level, we consider the use of non-parametric models, such as Gaussian Processes, to deal with both sensor mixtures and the variable scale of large outdoor environments. The talk is illustrated by a number of air, ground and sub-sea applications.
Stereoscopic vision is a passive solution to the problem of 3D imaging. This talk begins by reviewing the stereoscopic matching problem with in terms of 2D Markov Random Fields. Perhaps the most convincing solutions yet to the resulting 2D optimization problems are the ones that use various forms of ?graph cut? algorithm. Then the Panum Proxy matching algorithm will be explained, motivated by the physical constraints of the human stereo vision system. The result is an algorithm that is rather more efficient than usual for high resolution image pairs. Finally a few recent variants on stereo matching will be described and an interesting application that may help save the planet.
Over the last decade, the robotics community has developed highly efficient and robust solutions to state estimation problems such as robot localization, people tracking, and map building. With the availability of various techniques for spatially consistent sensor integration, an important next goal is to enable robots to reason about the many objects located in our everyday environments and to reason about spatial concepts such as rooms, hallways, streets, and intersections. An additional requirement for successful operation in populated environments is the ability to recognize the intent of humans and to adapt to their behavior patterns. In this talk I will present an overview of some recent work using graphical models and machine learning techniques to extract high level information from raw sensor data. Examples include place and object recognition from vision and laser data, and human activity recognition from wearable sensor data. I will conclude with an outlook for research directions in autonomous systems.
Over the last years, there has been a tremendous progress in the area of mobile robot navigation. Robots are able to build large-scale maps of their environments and to use these maps for navigation. However, the final step towards fully autonomous robots in real-world scenarios and industrial applications has not entirely been taken. In this presentation, I will describe some state-of-the-art techniques for robot navigation, potentials, and open research questions for taking the leap towards truly autonomous robots operating over long periods of time in complex and dynamic application scenarios.
Symposium Program Committee
Peter Druschel, MPI for Software Systems (chair)
Bernhard Schoelkopf, MPI for Biological Cybernetics
Roland Siegwart, Swiss Federal Institute of Technology, Zurich
Sebastian Thrun, Stanford University
Phone (Reception): +49 7071 601 551
Fax: +49 7071 601 552
How to reach the MPI campus
Please follow this link to get detailed information and maps how you reach the MPI campus.