Stefan Zickler

SSL-vision: the shared vision system for the robocup small size league

The current RoboCup Small Size League rules allow every team to set up their own global vision system as a primary sensor. This option, which is used by all participating teams, bears several organizational limitations and thus impairs the league’s progress. Additionally, most teams have converged on very similar solutions, and have produced only few significant research results to this global vision problem over the last years. Hence the responsible committees decided to migrate to a shared vision system (including also sharing the vision hardware) for all teams by 2010. This system – named SSL-Vision – is currently developed by volunteers from participating teams. In this paper, we describe the current state of SSL-Vision, i.e. its software architecture as well as the approaches used for image processing and camera calibration, together with the intended process for its introduction and its use beyond the scope of the Small Size League.

Sparselet models for efficient multiclass object detection

We develop an intermediate representation for deformable part models and show that this representation has favorable performance characteristics for multi-class problems when the number of classes is high. Our model uses sparse coding of part filters to represent each filter as a sparse linear combination of shared dictionary elements. This leads to a universal set of parts that are shared among all object classes. Reconstruction of the original part filter responses via sparse matrix-vector product reduces computation relative to conventional part filter convolutions. Our model is well suited to a parallel implementation, and we report a new GPU DPM implementation that takes advantage of sparse coding of part filters. The speed-up offered by our intermediate representation and parallel computation enable real-time DPM detection of 20 different object classes on a laptop computer.

Detection and Localization of Multiple Objects

Being able to identify and localize objects is an important requirement for various humanoid robot applications. In this paper we present a method which uses PCA-SIFT in combination with a clustered voting scheme to achieve detection and localization of multiple objects in real-time video data. Our approach provides robustness against constraints that are common for humanoid vision systems such as perspective changes, partial occlusion, and motion blurring. We analyze and evaluate the performance of our method in two concrete humanoid test-scenarios

RSS-based relative localization and tethering for moving robots in unknown environments

The LANdroids project requires robots to autonomously localize, track, and follow (a task also known as tethering) other robots or humans in an unknown environment with limited sensing abilities. In this paper, we present a localization and tethering approach that relies solely on wireless signal strength and robot odometry without requiring any known reference points in the domain. We introduce a data-driven, probabilistic model that maps received signal strength (RSS) values to real-world distance distributions and embed this model in a grid-based localization algorithm that successfully performs the LANdroids tethering task. We furthermore show, that it is possible to improve localization through the addition of a compass sensor and inter-robot information sharing.

Analyzing Multi-agent Activity Logs Using Process Mining Techniques

Distributed autonomous robotic systems exhibit complex behavior that-although programmed, but due to the impact of the environment- only materializes as the process unfolds. Thus, the actual behavior of such a system cannot be known in advance but must be observed to be evaluated or verified. In this paper we propose to use process mining techniques to extract, compare, and enhance models of the actual behavior of a multi-agent robotic system through analyzing collected log data. We use the example of robot soccer as such a multi-agent robotic system, and we demonstrate which types of analysis are currently possible in the context of the process mining tool set ProM.

Variable Level-Of-Detail Motion Planning in Environments with Poorly Predictable Bodies

Motion planning in dynamic environments consists of the generation of a collision-free trajectory from an initial to a goal state. When the environment contains uncertainty, preventing a perfect predictive model of its dynamics, a robot ends up only successfully executing a short part of the plan and then requires replanning, using the latest observed state of the environment. Each such replanning step is computationally expensive. Furthermore, we note that such sophisticated planning effort is unnecessary as the resulting plans are not likely to ever be fully executed, due to an unpredictable and changing environment. In this paper, we introduce the concept of Variable Level-Of-Detail (VLOD) planning, that is able to focus its search on obtaining accurate short-term results, while considering the far-future with a different level of detail, selectively ignoring the physical interactions with poorly predictable dynamic objects (e.g., other mobile bodies that are controlled by external entities). Unlike finite-horizon planning, which limits the maximum search depth, VLOD planning deals with local minima and generates full plans to the goal, while requiring much less computation than traditional planning. We contribute VLOD planning on a rich simulated physics-based planner and show results for varying LOD thresholds and replanning intervals.

CMDragons: Dynamic passing and strategy on a champion robot soccer team

After several years of developing multiple RoboCup small-size robot soccer teams, our CMDragons robot team achieved a highly successful level of performance, winning both the 2006 and 2007 competitions without losing a single game. Our small-size team consists of five executing wheeled robots with centralized, off-board perception and decision making. The decision making framework consists of a set of layered components, consisting of perception, evaluation and strategy, robot tactics and skills, and real-time navigation. In this paper, we present the strategy, action selection, and execution aspects of our architecture, with a focus on passing as an example of effective coordinated teamwork. The design enabled our robot team to score using multiple methods, from direct shooting up to 3D passes deflected in midair, resulting in a rich set of actions that were difficult for adversaries to counter. We provide several performance quantified claims supported by testing in our laboratory and in competition settings.

Generalized Sparselet Models for Real-Time Multiclass Object Recognition

The problem of real-time multiclass object recognition is of great practical importance in object recognition. In this paper, we describe a framework that simultaneously utilizes shared representation, reconstruction sparsity, and parallelism to enable real-time multiclass object detection with deformable part models at 5Hz on a laptop computer with almost no decrease in task performance. Our framework is trained in the standard structured output prediction formulation and is generically applicable for speeding up object recognition systems where the computational bottleneck is in multiclass, multi-convolutional inference. We experimentally demonstrate the efficiency and task performance of our method on PASCAL VOC, subset of ImageNet, Caltech101 and Caltech256 dataset.

Playing Creative Soccer: Randomized Behavioral Kinodynamic Planning of Robot Tactics

Modern robot soccer control architectures tend to separate higher level tactics and lower level navigation control. This can lead to tactics which do not fully utilize the robots dynamic actuation abilities. It can furthermore create the problem of the navigational code breaking the constraints of the higher level tactical goals when avoiding obstacles. We aim to improve such control architectures by modeling tactics as sampling-based behaviors which exist inside of a probabilistic kinodynamic planner, thus treating tactics and navigation as a unified dynamics problem. We present a behavioral version of Kinodynamic Rapidly- Exploring Random Trees and show that this algorithm can be used to automatically improvise new ball-manipulation strategies in a simulated robot soccer domain. We furthermore show how opponent-models can be seamlessly integrated into the planner, thus allowing the robot to anticipate and outperform the opponents motions in physics-space.

Depth-augmented Deformable Parts Models for RGBD person detection on embedded GPUs

Accurate real-time person detection is an important capability for many robot tasks, such as indoor navigation and human-robot interaction. In this paper, we introduce a depth-augmented, GPU-accelerated version of Deformable Parts Models (DPM) that uses a joint RGB+Depth feature descriptor to perform high-accuracy person detection at 5Hz while requiring less than 10 Watts on a single 2014 consumer-grade embedded chip. We provide a detailed description of the algorithm and evaluate its speed/accuracy trade-offs on an indoor person detection dataset collected from a mobile platform, showing that our RGBD approach outperforms accuracy of RGB-only DPM, depth-only DPM, and RGB HOG SVM classifier cascades. We furthermore demonstrate how reductions in model complexity and feature space dimensionality can increase speed without significantly sacrificing detector accuracy.