Refine
Departments, institutes and facilities
Document Type
- Conference Object (72)
- Preprint (15)
- Article (11)
- Report (5)
- Part of a Book (3)
- Book (monograph, edited volume) (1)
- Research Data (1)
- Doctoral Thesis (1)
Year of publication
Keywords
- Automatic Short Answer Grading (2)
- Cognitive robot control (2)
- Drosophila (2)
- Explainable robotics (2)
- Learning from experience (2)
- Navigation (2)
- Object detection (2)
- Object recognition (2)
- robot execution failures (2)
- robotics (2)
Robot Action Diagnosis and Experience Correction by Falsifying Parameterised Execution Models
(2021)
When faced with an execution failure, an intelligent robot should be able to identify the likely reasons for the failure and adapt its execution policy accordingly. This paper addresses the question of how to utilise knowledge about the execution process, expressed in terms of learned constraints, in order to direct the diagnosis and experience acquisition process. In particular, we present two methods for creating a synergy between failure diagnosis and execution model learning. We first propose a method for diagnosing execution failures of parameterised action execution models, which searches for action parameters that violate a learned precondition model. We then develop a strategy that uses the results of the diagnosis process for generating synthetic data that are more likely to lead to successful execution, thereby increasing the set of available experiences to learn from. The diagnosis and experience correction methods are evaluated for the problem of handle grasping, such that we experimentally demonstrate the effectiveness of the diagnosis algorithm and show that corrected failed experiences can contribute towards improving the execution success of a robot.
The majority of biomedical knowledge is stored in structured databases or as unstructured text in scientific publications. This vast amount of information has led to numerous machine learning-based biological applications using either text through natural language processing (NLP) or structured data through knowledge graph embedding models (KGEMs). However, representations based on a single modality are inherently limited. To generate better representations of biological knowledge, we propose STonKGs, a Sophisticated Transformer trained on biomedical text and Knowledge Graphs. This multimodal Transformer uses combined input sequences of structured information from KGs and unstructured text data from biomedical literature to learn joint representations. First, we pre-trained STonKGs on a knowledge base assembled by the Integrated Network and Dynamical Reasoning Assembler (INDRA) consisting of millions of text-triple pairs extracted from biomedical literature by multiple NLP systems. Then, we benchmarked STonKGs against two baseline models trained on either one of the modalities (i.e., text or KG) across eight different classification tasks, each corresponding to a different biological application. Our results demonstrate that STonKGs outperforms both baselines, especially on the more challenging tasks with respect to the number of classes, improving upon the F1-score of the best baseline by up to 0.083. Additionally, our pre-trained model as well as the model architecture can be adapted to various other transfer learning applications. Finally, the source code and pre-trained STonKGs models are available at https://github.com/stonkgs/stonkgs and https://huggingface.co/stonkgs/stonkgs-150k.
When an autonomous robot learns how to execute actions, it is of interest to know if and when the execution policy can be generalised to variations of the learning scenarios. This can inform the robot about the necessity of additional learning, as using incomplete or unsuitable policies can lead to execution failures. Generalisation is particularly relevant when a robot has to deal with a large variety of objects and in different contexts. In this paper, we propose and analyse a strategy for generalising parameterised execution models of manipulation actions over different objects based on an object ontology. In particular, a robot transfers a known execution model to objects of related classes according to the ontology, but only if there is no other evidence that the model may be unsuitable. This allows using ontological knowledge as prior information that is then refined by the robot’s own experiences. We verify our algorithm for two actions – grasping and stowing everyday objects – such that we show that the robot can deduce cases in which an existing policy can generalise to other objects and when additional execution knowledge has to be acquired.
Swedish wheeled mobile robots have remarkable mobility properties allowing them to rotate and translate at the same time. Being holonomic systems, their kinematics model results in the possibility of designing separate and independent position and heading trajectory tracking control laws. Nevertheless, if these control laws should be implemented in the presence of unaccounted actuator saturation, the resulting saturated linear and angular velocity commands could interfere with each other thus dramatically affecting the overall expected performance. Based on Lyapunov’s direct method, a position and heading trajectory tracking control law for Swedish wheeled robots is developed. It explicitly accounts for actuator saturation by using ideas from a prioritized task based control framework.
Co-design is concerned with the joint design of hardware and software making up an embedded computer system [Wol94]. A top down design flow for an embedded system begins with a system specification. If it is executable, it may be used for simulation, system verification or to identify algorithmical bottlenecks. In contrast to other chapters of this book, the specification is not developed in this case study, rather it is given from the beginning. Furthermore we are not concerned with partitioning or synthesis of dedicated HW. Instead we focus on the problem how to find an off-the-shelf micro-controller which implements the desired functionality and meets all specification constraints. If feasible, this is usually much cheaper then using dedicated hardware. This chapter will answer the question of feasibility for a real life problem from automobile industry.
Background: Virtual reality combined with spherical treadmills is used across species for studying neural circuits underlying navigation.
New Method: We developed an optical flow-based method for tracking treadmil ball motion in real-time using a single high-resolution camera.
Results: Tracking accuracy and timing were determined using calibration data. Ball tracking was performed at 500 Hz and integrated with an open source game engine for virtual reality projection. The projection was updated at 120 Hz with a latency with respect to ball motion of 30 ± 8 ms.
Comparison: with Existing Method(s) Optical flow based tracking of treadmill motion is typically achieved using optical mice. The camera-based optical flow tracking system developed here is based on off-the-shelf components and offers control over the image acquisition and processing parameters. This results in flexibility with respect to tracking conditions – such as ball surface texture, lighting conditions, or ball size – as well as camera alignment and calibration.
Conclusions: A fast system for rotational ball motion tracking suitable for virtual reality animal behavior across different scales was developed and characterized.
With regard to performance well established SW-only design methodologies proceed by making the initial specification run first, then by enhancing its functionality and finally by optimizing it. When designing Embedded Systems (EbS) this approach is not viable since decisive design decisions like e.g. the estimation of required processing power or the identification of those parts of the specification which need to be delegated to dedicated HW depend on the fastness and fairness of the initial specification. We here propose a sequence of optimization steps embedded into the design flow, which enables a structured way to accelerate a given working EbS specification at different layers. This sequence of accelerations comprises algorithm selection, algorithm transformation, data transformation, implementation optimization and finally HW acceleration. It is analyzed how all acceleration steps are influenced by the specific attributes of the underlying EbS. The overall acceleration procedure is explained and quantified at hand of a real-life industrial example.
SISAL: User manual
(1990)
Cognitive robotics aims at understanding biological processes, though it has also the potential to improve future robotics systems. Here we show how a biologically inspired model of motor control with neural fields can be augmented with additional components such that it is able to solve a basic robotics task, that of obstacle avoidance. While obstacle avoidance is a well researched area, the focus here is on the extensibility of a biologically inspired framework. This work demonstrates how easily the biological inspired system can be used to adapt to new tasks. This flexibility is thought to be a major hallmark of biological agents.
Application of underwater robots are on the rise, most of them are dependent on sonar for underwater vision, but the lack of strong perception capabilities limits them in this task. An important issue in sonar perception is matching image patches, which can enable other techniques like localization, change detection, and mapping. There is a rich literature for this problem in color images, but for acoustic images, it is lacking, due to the physics that produce these images. In this paper we improve on our previous results for this problem (Valdenegro-Toro et al, 2017), instead of modeling features manually, a Convolutional Neural Network (CNN) learns a similarity function and predicts if two input sonar images are similar or not. With the objective of improving the sonar image matching problem further, three state of the art CNN architectures are evaluated on the Marine Debris dataset, namely DenseNet, and VGG, with a siamese or two-channel architecture, and contrastive loss. To ensure a fair evaluation of each network, thorough hyper-parameter optimization is executed. We find that the best performing models are DenseNet Two-Channel network with 0.955 AUC, VGG-Siamese with contrastive loss at 0.949 AUC and DenseNet Siamese with 0.921 AUC. By ensembling the top performing DenseNet two-channel and DenseNet-Siamese models overall highest prediction accuracy obtained is 0.978 AUC, showing a large improvement over the 0.91 AUC in the state of the art.
CASTLE is a co-design platform developed at GMD SET institute. It provides a number of design tools for configuring application specific design flows. This paper presents a walk through the CASTLE co-design environment, following the design flow of a video processing system. The design methodology and the tool usage for this real life example are described, as seen from a designers point of view. The design flow starts with a C/C++ program and gradually derives a register-transfer level description of a processor hardware, as well as the corresponding compiler for generating the processor opcode. The main results of each design step are presented and the usage of the CASTLE tools at each step is explained.
Loading of shipping containers for dairy products often includes a press-fit task, which involves manually stacking milk cartons in a container without using pallets or packaging. Automating this task with a mobile manipulator can reduce worker strain, and also enhance the efficiency and safety of the container loading process. This paper proposes an approach called Adaptive Compliant Control with Integrated Failure Recovery (ACCIFR), which enables a mobile manipulator to reliably perform the press-fit task. We base the approach on a demonstration learning-based compliant control framework, such that we integrate a monitoring and failure recovery mechanism for successful task execution. Concretely, we monitor the execution through distance and force feedback, detect collisions while the robot is performing the press-fit task, and use wrench measurements to classify the direction of collision; this information informs the subsequent recovery process. We evaluate the method on a miniature container setup, considering variations in the (i) starting position of the end effector, (ii) goal configuration, and (iii) object grasping position. The results demonstrate that the proposed approach outperforms the baseline demonstration-based learning framework regarding adaptability to environmental variations and the ability to recover from collision failures, making it a promising solution for practical press-fit applications.
In the design of robot skills, the focus generally lies on increasing the flexibility and reliability of the robot execution process; however, typical skill representations are not designed for analysing execution failures if they occur or for explicitly learning from failures. In this paper, we describe a learning-based hybrid representation for skill parameterisation called an execution model, which considers execution failures to be a natural part of the execution process. We then (i) demonstrate how execution contexts can be included in execution models, (ii) introduce a technique for generalising models between object categories by combining generalisation attempts performed by a robot with knowledge about object similarities represented in an ontology, and (iii) describe a procedure that uses an execution model for identifying a likely hypothesis of a parameterisation failure. The feasibility of the proposed methods is evaluated in multiple experiments performed with a physical robot in the context of handle grasping, object grasping, and object pulling. The experimental results suggest that execution models contribute towards avoiding execution failures, but also represent a first step towards more introspective robots that are able to analyse some of their execution failures in an explicit manner.
Saliency methods are frequently used to explain Deep Neural Network-based models. Adebayo et al.'s work on evaluating saliency methods for classification models illustrate certain explanation methods fail the model and data randomization tests. However, on extending the tests for various state of the art object detectors we illustrate that the ability to explain a model is more dependent on the model itself than the explanation method. We perform sanity checks for object detection and define new qualitative criteria to evaluate the saliency explanations, both for object classification and bounding box decisions, using Guided Backpropagation, Integrated Gradients, and their Smoothgrad versions, together with Faster R-CNN, SSD, and EfficientDet-D0, trained on COCO. In addition, the sensitivity of the explanation method to model parameters and data labels varies class-wise motivating to perform the sanity checks for each class. We find that EfficientDet-D0 is the most interpretable method independent of the saliency method, which passes the sanity checks with little problems.
The dataset contains the following data from successful and failed executions of the Toyota HSR robot placing a book on a shelf.
RGB images from the robot's head camera
Depth images from the robot's head camera
Rendered images of the robot's 3D model from the point of view of the robot's head camera
Force-torque readings from a wrist-mounted force-torque sensor
Joint efforts, velocities and positions
extrinsic and intrinsic camera calibration parameters
frame-level anomaly annotations
The anomalies that occur during execution include:
the manipulated book falling down
books on the shelf being disturbed significantly
camera occlusions
robot being disturbed by an external collision
The dataset is split into a train, validation and test set with the following number of trials:
Train: 48 successful trials
Validation: 6 successful trials
Test: 60 anomalous trials and 7 successful trials
Data-Driven Robot Fault Detection and Diagnosis Using Generative Models: A Modified SFDD Algorithm
(2019)
This paper presents a modification of the data-driven sensor-based fault detection and diagnosis (SFDD) algorithm for online robot monitoring. Our version of the algorithm uses a collection of generative models, in particular restricted Boltzmann machines, each of which represents the distribution of sliding window correlations between a pair of correlated measurements. We use such models in a residual generation scheme, where high residuals generate conflict sets that are then used in a subsequent diagnosis step. As a proof of concept, the framework is evaluated on a mobile logistics robot for the problem of recognising disconnected wheels, such that the evaluation demonstrates the feasibility of the framework (on the faulty data set, the models obtained 88.6% precision and 75.6% recall rates), but also shows that the monitoring results are influenced by the choice of distribution model and the model parameters as a whole.
Tell Your Robot What To Do: Evaluation of Natural Language Models for Robot Command Processing
(2019)
The use of natural language to indicate robot tasks is a convenient way to command robots. As a result, several models and approaches capable of understanding robot commands have been developed, which however complicates the choice of a suitable model for a given scenario. In this work, we present a comparative analysis and benchmarking of four natural language understanding models - Mbot, Rasa, LU4R, and ECG. We particularly evaluate the performance of the models to understand domestic service robot commands by recognizing the actions and any complementary information in them in three use cases: the RoboCup@Home General Purpose Service Robot (GPSR) category 1 contest, GPSR category 2, and hospital logistics in the context of the ROPOD project.
In Sensor-based Fault Detection and Diagnosis (SFDD) methods, spatial and temporal dependencies among the sensor signals can be modeled to detect faults in the sensors, if the defined dependencies change over time. In this work, we model Granger causal relationships between pairs of sensor data streams to detect changes in their dependencies. We compare the method on simulated signals with the Pearson correlation, and show that the method elegantly handles noise and lags in the signals and provides appreciable dependency detection. We further evaluate the method using sensor data from a mobile robot by injecting both internal and external faults during operation of the robot. The results show that the method is able to detect changes in the system when faults are injected, but is also prone to detecting false positives. This suggests that this method can be used as a weak detection of faults, but other methods, such as the use of a structural model, are required to reliably detect and diagnose faults.
Cosynthesis in CASTLE
(1995)
Dual Dynamics (DD) is a mathematical model of a behavior control system for mobile autonomous robots. Behaviors are specified through differential equations; forming a global dynamical system made of behavior subsystems which interact in a number of ways. DD models can be directly compiled into executable code. The article (i) explains the model; (ii) sketches the Dual Dynamics Designer (DDD) environment that we use for the design; simulation; implementation and documentation; and (iii) illustrates our approach with the example of kicking a moving ball into a goal.
A way of combining a relatively new sensor-technology, that is optical analog VLSI devices, with a standard digital omni-directional vision system is investigated. The sensor used is a neuromorphic analog VLSI sensor that estimates the global visual image motion. The sensor provides two analog output voltages that represent the components of the global optical flow vector. The readout is guided by an omni-directional mirror that maps the location of the ball and directs the robot to align its position so that a sensor-actuator module that includes the analog VLSI optical flow sensor can be activated. The purpose of the sensor-actuator module is to operate with a higher update rate than the standard vision system and thus increase the reactivity of the robot for very specific situations. This paper will demonstrate an application example where the robot is a goalkeeper with the task of defending the goal during a penalty kick.
Ein gebräuchliche Methodik beim Entwurf eingebetteter Systeme, in Anwendung besonders bei kleinen- und mittleren Unternehmen, geht folgendermaßen vor: Man nehme das bereits existierende Mikrokontroller Entwicklungspaket und bereits vorhandene Funktionen aus einer alten Systemrealisierung, variiere bzw. passe sie an die neue Aufgabe an und teste dann durch Emulation, ob die Spezifikation erfüllt ist.
GMD-Robots
(2001)
The ability to track moving people is a key aspect of autonomous robot systems in real-world environments. Whilst for many tasks knowing the approximate positions of people may be sufficient, the ability to identify unique people is needed to accurately count people in the real world. To accomplish the people counting task, a robust system for people detection, tracking and identification is needed.
An essential measure of autonomy in assistive service robots is adaptivity to the various contexts of human-oriented tasks, which are subject to subtle variations in task parameters that determine optimal behaviour. In this work, we propose an apprenticeship learning approach to achieving context-aware action generalization on the task of robot-to-human object hand-over. The procedure combines learning from demonstration and reinforcement learning: a robot first imitates a demonstrator’s execution of the task and then learns contextualized variants of the demonstrated action through experience. We use dynamic movement primitives as compact motion representations, and a model-based C-REPS algorithm for learning policies that can specify hand-over position, conditioned on context variables. Policies are learned using simulated task executions, before transferring them to the robot and evaluating emergent behaviours. We additionally conduct a user study involving participants assuming different postures and receiving an object from a robot, which executes hand-overs by either imitating a demonstrated motion, or adapting its motion to hand-over positions suggested by the learned policy. The results confirm the hypothesized improvements in the robot’s perceived behaviour when it is context-aware and adaptive, and provide useful insights that can inform future developments.
Deep learning models are extensively used in various safety critical applications. Hence these models along with being accurate need to be highly reliable. One way of achieving this is by quantifying uncertainty. Bayesian methods for UQ have been extensively studied for Deep Learning models applied on images but have been less explored for 3D modalities such as point clouds often used for Robots and Autonomous Systems. In this work, we evaluate three uncertainty quantification methods namely Deep Ensembles, MC-Dropout and MC-DropConnect on the DarkNet21Seg 3D semantic segmentation model and comprehensively analyze the impact of various parameters such as number of models in ensembles or forward passes, and drop probability values, on task performance and uncertainty estimate quality. We find that Deep Ensembles outperforms other methods in both performance and uncertainty metrics. Deep ensembles outperform other methods by a margin of 2.4% in terms of mIOU, 1.3% in terms of accuracy, while providing reliable uncertainty for decision making.
For robots acting - and failing - in everyday environments, a predictable behaviour representation is important so that it can be utilised for failure analysis, recovery, and subsequent improvement. Learning from demonstration combined with dynamic motion primitives is one commonly used technique for creating models that are easy to analyse and interpret; however, mobile manipulators complicate such models since they need the ability to synchronise arm and base motions for performing purposeful tasks. In this paper, we analyse dynamic motion primitives in the context of a mobile manipulator - a Toyota Human Support Robot (HSR)- and introduce a small extension of dynamic motion primitives that makes it possible to perform whole body motion with a mobile manipulator. We then present an extensive set of experiments in which our robot was grasping various everyday objects in a domestic environment, where a sequence of object detection, pose estimation, and manipulation was required for successfully completing the task. Our experiments demonstrate the feasibility of the proposed whole body motion framework for everyday object manipulation, but also illustrate the necessity for highly adaptive manipulation strategies that make better use of a robot's perceptual capabilities.
MOTIVATION
The majority of biomedical knowledge is stored in structured databases or as unstructured text in scientific publications. This vast amount of information has led to numerous machine learning-based biological applications using either text through natural language processing (NLP) or structured data through knowledge graph embedding models (KGEMs). However, representations based on a single modality are inherently limited.
RESULTS
To generate better representations of biological knowledge, we propose STonKGs, a Sophisticated Transformer trained on biomedical text and Knowledge Graphs (KGs). This multimodal Transformer uses combined input sequences of structured information from KGs and unstructured text data from biomedical literature to learn joint representations in a shared embedding space. First, we pre-trained STonKGs on a knowledge base assembled by the Integrated Network and Dynamical Reasoning Assembler (INDRA) consisting of millions of text-triple pairs extracted from biomedical literature by multiple NLP systems. Then, we benchmarked STonKGs against three baseline models trained on either one of the modalities (i.e., text or KG) across eight different classification tasks, each corresponding to a different biological application. Our results demonstrate that STonKGs outperforms both baselines, especially on the more challenging tasks with respect to the number of classes, improving upon the F1-score of the best baseline by up to 0.084 (i.e., from 0.881 to 0.965). Finally, our pre-trained model as well as the model architecture can be adapted to various other transfer learning applications.
AVAILABILITY
We make the source code and the Python package of STonKGs available at GitHub (https://github.com/stonkgs/stonkgs) and PyPI (https://pypi.org/project/stonkgs/). The pre-trained STonKGs models and the task-specific classification models are respectively available at https://huggingface.co/stonkgs/stonkgs-150k and https://zenodo.org/communities/stonkgs.
SUPPLEMENTARY INFORMATION
Supplementary data are available at Bioinformatics online.
Property-Based Testing in Simulation for Verifying Robot Action Execution in Tabletop Manipulation
(2021)
An important prerequisite for the reliability and robustness of a service robot is ensuring the robot’s correct behavior when it performs various tasks of interest. Extensive testing is one established approach for ensuring behavioural correctness; this becomes even more important with the integration of learning-based methods into robot software architectures, as there are often no theoretical guarantees about the performance of such methods in varying scenarios. In this paper, we aim towards evaluating the correctness of robot behaviors in tabletop manipulation through automatic generation of simulated test scenarios in which a robot assesses its performance using property-based testing. In particular, key properties of interest for various robot actions are encoded in an action ontology and are then verified and validated within a simulated environment. We evaluate our framework with a Toyota Human Support Robot (HSR) which is tested in a Gazebo simulation. We show that our framework can correctly and consistently identify various failed actions in a variety of randomised tabletop manipulation scenarios, in addition to providing deeper insights into the type and location of failures for each designed property.
Execution monitoring is essential for robots to detect and respond to failures. Since it is impossible to enumerate all failures for a given task, we learn from successful executions of the task to detect visual anomalies during runtime. Our method learns to predict the motions that occur during the nominal execution of a task, including camera and robot body motion. A probabilistic U-Net architecture is used to learn to predict optical flow, and the robot's kinematics and 3D model are used to model camera and body motion. The errors between the observed and predicted motion are used to calculate an anomaly score. We evaluate our method on a dataset of a robot placing a book on a shelf, which includes anomalies such as falling books, camera occlusions, and robot disturbances. We find that modeling camera and body motion, in addition to the learning-based optical flow prediction, results in an improvement of the area under the receiver operating characteristic curve from 0.752 to 0.804, and the area under the precision-recall curve from 0.467 to 0.549.
This paper presents an approach to estimate theego-motion of a robot while moving. The employed sensor is aTime-of-Flight (ToF) camera, the SR3000 from Mesa Imaging.ToF cameras provide depth and reflectance data of the scene athigh frame rates.The proposed method utilizes the coherence of depth andreflectance data of ToF cameras by detecting image features onreflectance data and estimating the motion on depth data. Themotion estimate of the camera is fused with inertial measure-ments to gain higher accuracy and robustness.The result of the algorithm is benchmarked against referenceposes determined by matching accurate 2D range scans. Theevaluation shows that fusing the pose estimate with the datafromthe IMU improves the accuracy and robustness of the motionestimate against distorted measurements from the sensor.