previous day
all days

View: session overviewtalk overview

09:00-10:30 Session 14: Robots and Agents I
INVITED TALK: Digital Twin Knowledge Bases -- Knowledge Representation and Reasoning for Robotic Agents

ABSTRACT. Robotic agents that can accomplish manipulation tasks with the
competence of humans have been one of the grand research challenges
for artificial intelligence (AI) and robotics research for more than
50 years. However, while the fields made huge progress over the years,
this ultimate goal is still out of reach. I believe that this is the
case because the knowledge representation and reasoning methods that
have been proposed in AI so far are necessary but too abstract. In
this talk I propose to address this problem by endowing robots with
the capability to internally emulate and simulate their
perception-action loops based on realistic images and faithful physics
simulations, which are made machine-understandable by casting them as
virtual symbolic knowledge bases. These capabilities allow robots to
generate huge collections of machine-understandable manipulation
experiences, which robotic agents can generalize into commonsense and
intuitive physics knowledge applicable to open varieties of
manipulation tasks. The combination of learning, representation, and
reasoning will equip robots with an understanding of the relation
between their motions and the physical effects they cause at an
unprecedented level of realism, depth, and breadth, and enable them to
master human-scale manipulation tasks. This breakthrough will be
achievable by combining leading-edge simulation and visual rendering
technologies with mechanisms to semantically interpret and introspect
internal simulation data structures and processes. Robots with such
manipulation capabilities can help us to better deal with important
societal, humanitarian, and economic challenges of our aging

Towards Intuitive Robot Programming Using Finite State Automata

ABSTRACT. This paper describes an approach to intuitive robot programming, with the aim of enabling non-experts to generate sensor-based, structured programs. The core idea is to generate a variant of a finite state automaton (representing the program) by kinesthetic programming (physically guiding the robot). We use the structure of the automaton for control flow (loops and branching according to conditions of the environment). For programming, we forgo a visual user interface completely to determine to what extent this is viable. Our experiments show that non-expert users are indeed able to successfully program small sample tasks within reasonable time.

ALICA 2.0 - Domain-Independent Teamwork

ABSTRACT. We present a new version of ALICA - "A Language for Interactive Cooperative Agents". The ALICA framework is a highly reactive multi-agent framework and comprises three components for working with multi-agent plans: a specification language, an execution engine, and a graphical modelling tool. The framework automatically coordinates teams, allocates tasks to team members, and compensates execution failures in a fully distributed manner. In a major redesign, we extended the description language and re-implemented the execution engine and graphical modelling tool. As a result, the second version of ALICA encompasses fewer dependencies, is domain independent, and adaptable to different environments.

10:30-11:00Coffee Break
11:00-12:30 Session 15: Robots and Agents II
Automated Robot Skill Learning from Demonstration for Various Robot Systems

ABSTRACT. Transferring human movements to robotic systems is of high interest to equip the systems with new behaviors without expert knowledge. Typically, skills are often only learned for a very specific setup and a certain robot. We propose a modular framework to learn skills and is applicable on different robotic systems without adaptations. Our work builds on the recently introduced BesMan Learning Platform, which comprises the full workflow to transfer human demonstrations to a system, including automatized behavior segmentation, imitation learning, reinforcement learning for motion refinement and methods to generalize to related tasks. For this paper, we extend this approach in order that different skills can be imitated by various systems in an automated fashion with a minimal amount of configuration, e.g., definition of the target system and environment. For this, we focus on the imitation of the demonstrated movements and show their transferability without movement refinement. We demonstrate the generality of the approach on a large dataset, consisting of about 700 throwing demonstrations. Nearly all of these human demonstrations are successfully transferred to four different robot target systems, namely the Universal Robot's UR5 and UR10, the KUKA LBR iiwa and DFKI's robot COMPI. An analysis of the quality of the imitated movement on the real UR5 robot shows that useful throws can be executed on the system, that can be used as starting points for further movement refinement.

Epistemic Multi-Agent Planning Using Monte-Carlo Tree Search
PRESENTER: Daniel Reifsteck

ABSTRACT. Coordination in multi-agent systems with partial and non-uniform observability is a problem which has to be solved efficiently in practice. We use Monte-Carlo tree search as the basis of an implicitly coordinated epistemic planning algorithm which is capable of using the different knowledge of the agents to find solutions in problems despite a large branching factor. We use Dynamic Epistemic Logic to represent the knowledge and the actual situation as a state of the Monte-Carlo tree search, and epistemic planning to formalize the goals and actions of a problem. Further, we describe the required modifications of the Monte-Carlo tree search to use epistemic states, and make use of the cooperative card game Hanabi to test our planner on larger problems. We find that the approach scales to games with up to eight cards while maintaining high playing strength.

Monte-Carlo Search for Prize-Collecting Robot Motion Planning with Time Windows, Capacities, Pickups, and Deliveries

ABSTRACT. Logistics operations often require a robot to pickup and deliver objects from multiple locations within certain time frames. This is a challenging task-and-motion planning problem as it intertwines logical and temporal constraints about the operations with geometric and differential constraints related to obstacle avoidance and robot dynamics. To address these challenges, this paper couples vehicle routing over a discrete abstraction with sampling-based motion planning. On the one hand, vehicle routing provides plans to effectively guide sampling-based motion planning as it explores the vast space of feasible motions. On the other hand, motion planning provides feasibility estimates which vehicle routing uses to refine its plans. This coupling makes it possible to extend the state-of-the-art in multi-goal motion planning by also incorporating capacities, pickups, and deliveries in addition to time windows. When not all pickups and deliveries can be completed in time, the approach seeks to minimize the violations and maximize the profit.

Determining Fuzzy Controllers from Qualitative Interviews for Travel Mode Choice in an Agent-based Mobility Simulation

ABSTRACT. Both qualitative and quantitative research are integral parts for the understanding of traffic systems, yet it can be difficult to formalize and execute qualitative research results in a technical simulation system in an understandable and flexible manner. This paper presents an approach to systematically construct fuzzy controllers from socio-scientific data for the application in an agent-based mobility simulation. A general fuzzy controller concept is presented and subsequently applied to statements about travel mode choice and common activities from semi-structured interviews on mobility behavior. It is shown that the controller concept can be used to determine both fuzzy rule base and the linguistic variables and terms from the interviews and that such a controller can be used successfully in an agent-based mobility simulation.

12:30-14:00Lunch Break
14:00-15:30 Session 16: Knowledge Representation and Reasoning
Monotone and Online Fair Division

ABSTRACT. We study a new but simple model for online fair division in which indivisible items arrive one-by-one and agents have monotone utilities over bundles of the items. We consider axiomatic properties of mechanisms for this model such as strategy-proofness, envy-freeness and Pareto efficiency. We prove a number of impossibility results that justify why we consider relaxations of the properties, as well as why we consider restricted preference domains on which good axiomatic properties can be achieved. We propose two mechanisms that have good axiomatic fairness properties on restricted but common preference domains.

Simplifying Automated Pattern Selection for Planning with Symbolic Pattern Databases

ABSTRACT. Pattern databases (PDBs) are memory-based abstraction heuristics that are constructed prior to the planning process which, if expressed symbolically, yield a very efficient representation. Recent work in the automatic generation of symbolic PDBs has established it as one of the most successful approaches for cost-optimal domain-independent planning. In this paper, we contribute two planners, both using bin-packing for its pattern selection. In the second one, we introduce a greedy selection algorithm called Partial-Gamer, which complements the heuristic given by bin-packing. We tested our approaches on the benchmarks of the last three International Planning Competitions, optimal track, getting very competitive results, with this simple and deterministic algorithm.

A human-oriented system for equational reasoning

ABSTRACT. We introduce a fully automatic system, implemented in the Lean theorem prover, that solves equality problems of everyday mathematics. Our overriding priority in devising the system is that it should construct proofs of equality in a way that is similar to that of humans. A second goal is that the methods it uses should be domain independent. The basic strategy of the system is to operate with a subtask stack: whenever there is no clear way of making progress towards the task at the top of the stack, the program finds a promising subtask, such as rewriting a subterm, and places that at the top of the stack instead. Heuristics guide the choice of promising subtasks and the rewriting process. We show that our system can prove equality theorems simply, without having to preselect or orient rewrite rules as in standard theorem provers, and without having to invoke heavy duty tools for performing simple reasoning.

Towards Leveraging Backdoors in Qualitative Constraint Networks

ABSTRACT. In this technical communication we introduce the notions of backbones and backdoors in the context of qualitative constraint networks. As motivation for the study of those structures, we argue that they can be used to define collaborative approaches among SAT, CP, and native tools, inspire novel decomposition and parallelization techniques, and lead to the development of adaptive constraint propagators that will have a better insight into the particularities of real-world datasets than what is possible today.

15:30-16:00Coffee Break
16:00-17:15 Session 17: Bayesian and Neural Networks
Gaussian Lifted Marginal Filtering

ABSTRACT. Recently, Lifted Marginal Filtering has been proposed, an efficient Bayesian filtering algorithm for stochastic systems consisting of multiple, (inter-)acting agents and objects (entities). The algorithm achieves its efficiency by performing inference jointly over groups of similar entities (i.e. their properties follow the same distribution). In this paper, we explore the case where there are no entities that are directly suitable for grouping. We propose to use methods from Gaussian mixture fitting to identify entity groups, such that the error imposed by grouping them (by approximating their properties by a distribution) is minimal. Furthermore, we show how Gaussian mixture merging methods can be used to prevent the number of groups from growing indefinitely over time. We evaluate our approach on an activity prediction task in an online multiplayer game. The results suggest that compared to the conventional approach, where all entities are handled individually, decrease in prediction accuracy is small, while inference runtime decreases significantly.

GAN Path Finder: Preliminary results
PRESENTER: Natalia Soboleva

ABSTRACT. 2D path planning in known, static environment is a well-known problem in AI and robotics. One of the common ways to solve it is to a) represent the environment as a grid composed of blocked and traversable cells, b) perform a heuristic search for a path on such a grid. At the same time 2D grid resembles much a digital image, thus an appealing idea comes to being - to treat the problem as an image generation problem and to solve it utilizing the recent advances in artificial neural networks and deep learning. In this work we make an attempt to apply a generative model as a path finder. We create a context aware generative adversarial net that generates a path image in response to context input, i.e. image of the grid-map with start and goal. We demonstrate empirically that the model can successfully handle low-dimensional input and solve previously unseen instances.

Reducing Search Space of Genetic Algorithms for Fast Black Box Attacks on Image Classifiers
PRESENTER: Lukas Hartmann

ABSTRACT. Recent research regarding the reliability of Deep Neural Networks (DNN) revealed that it is easy to produce images that are completely unrecognizable to humans, but DNNs recognize as classifiable objects with 99.99% confidence. The present study investigates the effect of search space reduction for Genetic Algorithms (GA) on their capability of purposefully fooling DNNs. Therefore, we introduce a GA with respective modifications that is able to fool neural networks trained to classify objects from well-known benchmark image data sets like GTSRB or MNIST. The developed GA is extended and thus capable of reducing the search space without changing its general behavior. Empirical results on MNIST indicate a significantly decreased number of generations needed to satisfy the targeted confidence of an MNIST image classifier (12 instead of 228 generations). Conducted experiments on GTSRB, a more challenging object classification scenario, show similar results. Therefore, fooling DNNs has found not only easily possible but can also be done very fast. Our study thus substantiates an already recognized, potential danger for DNN-based computer vision or object recognition applications.

InformatiCup Competition 2019: Fooling Traffic Sign Recognition

ABSTRACT. Neural networks are used more and more in critical areas such as autonomous driving. In such cases, their limitations might cause dangerous situations. Researchers were able to show that such limitations enable attacks on systems containing neural networks, which are even possible in real world scenarios. For example, a state-of-the-art network might misclassify modified traffic signs. Other researchers have shown that modern car assistants can easily be fooled to drive the car into the wrong lane on a street.

The InformatiCup is a collegiate computer science competition in Germany, Switzerland and Austria for all students, with tasks based on real world problems. This year's task is based on the abovementioned problem. To demonstrate this problem and to motivate students for experimenting with neural networks, participants were asked to generate fooling images for a traffic sign classifying neural network without having direct access to the network. The images should not be recognisable by humans as traffic signs, but be classified as such with a high confidence by the neural network.