Thomas Hellström

A Formalism for Learning from Demonstration

The paper describes and formalizes the concepts and assumptions involved in Learning from Demonstration (LFD), a common learning technique used in robotics. LFD-related concepts like goal, generalization, and repetition are here defined, analyzed, and put into context. Robot behaviors are described in terms of trajectories through information spaces and learning is formulated as mappings between some of these spaces. Finally, behavior primitives are introduced as one example of good bias in learning, dividing the learning process into the three stages of behavior segmentation, behavior recognition, and behavior coordination. The formalism is exemplified through a sequence learning task where a robot equipped with a gripper arm is to move objects to specific areas. The introduced concepts are illustrated with special focus on how bias of various kinds can be used to enable learning from a single demonstration, and how ambiguities in demonstrations can be identified and handled.

On the moral responsibility of military robots

This article discusses mechanisms and principles for assignment of moral responsibility to intelligent robots, with special focus on military robots. We introduce the concept autonomous power as a new concept, and use it to identify the type of robots that call for moral considerations. It is furthermore argued that autonomous power, and in particular the ability to learn, is decisive for assignment of moral responsibility to robots. As technological development will lead to robots with increasing autonomous power, we should be prepared for a future when people blame robots for their actions. It is important to, already today, investigate the mechanisms that control human behavior in this respect. The results may be used when designing future military robots, to control unwanted tendencies to assign responsibility to the robots. Independent of the responsibility issue, the moral quality of robots’ behavior should be seen as one of many performance measures by which we evaluate robots. How to design ethics based control systems should be carefully investigated already now. From a consequentialist view, it would indeed be highly immoral to develop robots capable of performing acts involving life and death, without including some kind of moral framework.

Autonomous forest vehicles: historic, envisioned, and state-of-the-art.

Abstract The feasibility of using autonomous forest vehicles (which can be regarded as logical developments in the ongoing automation of forest machines), the systems that could be applied in them, their potential advantages and limitations (in the foreseeable future) are considered in this paper. The goals were to analyze: 1) the factors influencing the degree of automation in logging; 2) the technical principles that can be applied to autonomous forest machines, and 3) the feasibility of developing an autonomous path-tracking forest vehicle. A type of vehicle that is believed to have considerable commercial potential is an autonomous forwarder. The degree of automation is influenced by increased productivity, the machine operator as a bottle-neck, cost reduction, and environmental aspects. Technical principles that can be applied to autonomous forest vehicles are satellite navigation, wheel odometry, laser scanner, and radar. A new path-tracking algorithm has been developed to reduce deviations from the desired path by utilizing the driver’s steering commands. The presented system has demonstrated both possibilities and difficulties associated with autonomous forest machines. A field study has shown that it is quite possible for them to learn and track a path previously demonstrated by an operator with an accuracy of 0.1 m on flat ground and also to detect and avoid unexpected obstacles. Although the forest machine safely avoids obstacles, the study shows that further research in the field of obstacle avoidance is needed to optimize performance and ensure safe operation in a real forest environment.

Development of an Autonomous Forest Machine for Path Tracking

In many respects traditional automation in the forest-machine industry hasreached an upper limit, since the driver already has to deal with an excess ofinformation and take too many decisions at a ...

Behavior recognition for Learning from Demonstration

Two methods for behavior recognition are presented and evaluated. Both methods are based on the dynamic temporal difference algorithm Predictive Sequence Learning (PSL) which has previously been proposed as a learning algorithm for robot control. One strength of the proposed recognition methods is that the model PSL builds to recognize behaviors is identical to that used for control, implying that the controller (inverse model) and the recognition algorithm (forward model) can be implemented as two aspects of the same model. The two proposed methods, PSLE-Comparison and PSLH-Comparison, are evaluated in a Learning from Demonstration setting, where each algorithm should recognize a known skill in a demonstration performed via teleoperation. PSLH-Comparison produced the smallest recognition error. The results indicate that PSLH-Comparison could be a suitable algorithm for integration in a hierarchical control system consistent with recent models of human perception and motor control.

Enhanced Algorithms for Estimating Tree Trunk Diameter Using 2D Laser Scanner

Accurate vehicle localization in forest environments is still an unresolved problem. Global navigation satellite systems (GNSS) have well known limitations in dense forest, and have to be combined with for instance laser based SLAM algorithms to provide satisfying accuracy. Such algorithms typically require accurate detection of trees, and estimation of tree center locations in laser data. Both these operations depend on accurate estimations of tree trunk diameter. Diameter estimations are important also for several other forestry automation and remote sensing applications. This paper evaluates several existing algorithms for diameter estimation using 2D laser scanner data. Enhanced algorithms, compensating for beam width and using multiple scans, were also developed and evaluated. The best existing algorithms overestimated tree trunk diameter by ca. 40%. Our enhanced algorithms, compensating for laser beam width, reduced this error to less than 12%.

Visual tree detection for autonomous navigation in forest environment

This paper describes a classification based tree detection method for autonomous navigation of forest vehicles in forest environment. Fusion of color, and texture cues has been used to segment the image into tree trunk and background objects. The segmentation of images into tree trunk and background objects is a challenging task due to high variations of illumination, effect of different color shades, non-homogeneous bark texture, shadows and foreshortening. To accomplish this, the approach has been to find the best combinations of color, and texture descriptors, and classification techniques. An additional task has been to estimate the distance between forest vehicle and the base of segmented trees using monocular vision. A simple heuristic distance measurement method is proposed that is based on pixel height and a reference width. The performance of various color and texture operators, and accuracy of classifiers has been evaluated using cross validation techniques.

Path tracking in forest terrain by an autonomous forwarder

Abstract Autonomous navigation in forest terrain, where operation paths are rarely straight or flat and obstacles are common, is challenging. This paper evaluates a system designed to autonomously follow previously demonstrated paths in a forest environment without loading/unloading timber, a pre-step in the development of fully autonomous forwarders. The system consisted of a forwarder equipped with a high-precision global positioning system to measure the vehicles heading and position. A gyro was used to compensate for the influence of the vehicles roll and pitch. On an ordinary clear-cut forest area with numerous stumps, the vehicle was able to follow two different tracks, three times each at a speed of 1 m s−1, with a mean path tracking error of 6 and 7 cm, respectively. The error never exceeded 35 cm, and in 90% of the observations it was less than 14 and 15 cm, respectively. This accuracy is well within the necessary tolerance for forestry operations. In fact, a human operator would probably have a hard time following the track more accurately. Hence, the developed systems function satisfactorily when using previously demonstrated paths. However, further research on planning new paths in unknown unstructured terrain and on loading/unloading is required before timber transports can be fully automated.

Learning High-Level Behaviors From Demonstration Through Semantic Networks

In this paper we present an approach for high-level behavior recognition and selection integrated with alow-level controller to help the robot to learn new skills from demonstrations. By means of S ...

Real-time path planning using a simulator-in-the-loop

This paper describes the development of a real-time path planner for off-road vehicles using a simulator. The general idea with the presented system is to extend a standard path-tracking algorithm with a simulator that, in real-time, tries to predict collisions in a window forward in time. If a collision is predicted, the vehicle is stopped and a path-search phase is initiated. Variants of the original path are generated and simulated until a feasible path is found. The real vehicle then continues, now tracking the replanned path.