In previous work, a multi-layered neural network trust model, dubbed NeuroTrust, was introduced. This trust model was also implemented in an autonomous vehicles convoy simulation, in which speed and gap distance depended on trust. It has been shown that, in time, through on-line reinforcement learning, this trust model produces better results for significant performance metrics in the respective autonomous vehicle convoy when compared to a baseline trust algorithm. In this paper, the NeuroTrust model is expanded to leverage the experience of multiple decision-making agents. A trust aggregation method is proposed for NeuroTrust and is simulated for multiple autonomous vehicle convoy scenarios. It is shown that the NeuroTrust model tends to optimize faster by leveraging each agent’s experience.

In this paper we propose a trust algorithm, dubbed NeuroTrust, based on a multi-layered neural network. Previous work introduced trust as a performance estimation algorithm between team members in multi-agent systems, to allow for behavior optimization of the team. The trust model was developed based on an Acceptance Observation History (AOH) and confirmation and tolerance parameters to control trust growth and decay. Further work proposed certain improvements, in an autonomous vehicles convoy scenario, by considering agent diversity and a non-linear relationship between trust and vehicle control. In this work we show a further optimization using a deep recurrent neural network. This multi-layered neural network delivers trust as a probability function estimation with AOH as a sliding window batch input. The neural network is pre-trained using supervised learning, to emulate the previous trust model, as baseline. This pre-trained model is then exposed to future optimization using on-line reinforcement learning. The proposed trust model could be adaptable to a variety of systems, external conditions, and agent diversity. One application example where such a biologically-inspired trust model is suitable would be for soldier-machine teaming. Furthermore, particularly in the autonomous convoy scenario, we can account for the trust-control relationship nonlinearity in the trust domain, thus simplifying the control algorithm.

In many multi-agent teams, entities fully trust their teammates and the information that they provide. But we know that this can be a false assumption in many cases, which can lead to sub-optimal performance of the team. In this paper, we build off of prior work in developing a simple model of estimating and responding to different levels of trust between team members. We have chosen to use a vehicle convoy application to generate data and test the operation of the trust estimation algorithm and its evolution. We build on prior work, where a cruise control algorithm to maintain following distance was implemented, as were algorithms to adjust follow distance based on trust in the leader and the capability for a lead vehicle to “look back” and adjust its speed based on the follow distance of the vehicle behind. In this paper we introduce a mechanism, based on trust, which negotiates between two follow behaviors, either follow the vehicle ahead or drive towards a set of fixed waypoints. We also add a nonlinear relationship between trust and follow distance to provide a knob to adjust convoy performance and the paper shows that it does adjust performance, somewhat as expected.

In many multi-agent systems, robots within the same team are regarded as being fully trustworthy for cooperative tasks. However, the assumption of trustworthiness is not always justified, which may not only increase the risk of mission failure, but also endanger the lives of friendly forces. In prior work, we addressed this issue by using RoboTrust to dynamically adjust to observed behaviors or recommendations in order to mitigate the risk of illegitimate behaviors. However, in those simulations, all members of the convoy had knowledge of the convoy goals. In this paper, only the lead vehicle has knowledge of the convoy goals and the follower vehicles must infer trustworthiness strictly from leader vehicle performance. RoboTrust also could only respond to observed performance and did not dynamically learn agent behavior. In this paper, however, we incorporate an adaptive agent-specific bias into the RoboTrust algorithm that modifies its trust dynamics. This bias is learned from agent interactions, allowing good agents to benefit from faster trust growth and slower trust decay and bad agents to be penalized with slower trust growth and faster trust decay. We then integrate this new trust model into a trust-based controller for decentralized autonomous convoy operations. We evaluate its performance in an obstacle avoidance mission, where the convoy attempts to learn the best speed and following distance combinations for an acceptable obstacle avoidance probability.

Proceedings of SPIE present the original research papers presented at SPIE conferences and other high-quality conferences in the broad-ranging fields of optics and photonics. These books provide prompt access to the latest innovations in research and technology in their respective fields. Proceedings of SPIE are among the most cited references in patent literature.

In the field of military Unmanned Ground Vehicles (UGV), military units are adapting their concept of operations to focus on their mission capabilities within populated cities and towns. These types of operations are referred to as MOUT (Military Operations on Urban Terrain). As more Soldiers seek to incorporate technology to enhance their mission capabilities, there then becomes a need for UGV systems to encompass an ability to autonomously navigate through urban terrains. Autonomous systems have the potential to increase Soldier safety by mitigating the risk of unnecessary enemy exposure during routine urban reconnaissance. This paper presents the development and methodology that the military has sought to increase mission capabilities by incorporating autonomy into manned/unmanned ground vehicles. The presented solution that has been developed through the Safe Operations of Unmanned systems for Reconnaissance in Complex Environments (SOURCE) Army Technology Objective (ATO) has the ability and has been tested to safely navigate through complex urban environments. This paper will also focus on the challenges the military has faced to develop the presented autonomous UGV.

As more Soldiers seek to utilize robots to enhance their mission capabilities, controls are needed which are intuitive, portable, and adaptable to a wide range of mission tasks. Android (TM) and iOS (TM) devices have the potential to meet each of these requirements as well as being based on readily available hardware. This paper will focus on some of the ways in which an Android (TM) or iOS (TM) device could be used to control specific and varied robot mobility functions and payload tools. Several small unmanned ground vehicle (SUGV) payload tools will have been investigated at Camp Pendleton during a user assessment and mission feasibility study for automatic remote tool changing. This group of payload tools will provide a basis, to researchers, concerning what types of control functions are needed to fully utilize SUGV robotic capabilities. Additional, mobility functions using tablet devices have been used as part of the Safe Operation of Unmanned systems for Reconnaissance in Complex Environments Army Technology Objective (SOURCE ATO) which is investigating the safe operation of robotics.Using Android (TM) and iOS (TM) hand-held devices is not a new concept in robot manipulation. However, the authors of this paper hope to introduce some novel concepts that may serve to make the interaction between Soldier and machine more fluid and intuitive. By creating a better user experience, Android (TM) and iOS (TM) devices could help to reduce training time, enhance performance, and increase acceptance of robotics as valuable mission tools for Soldiers.

The real world is too complex and variable to directly program an autonomous ground robot's control system to respond to the inputs from its environmental sensors such as LIDAR and video. The need for learning incrementally, discarding prior data, is important because of the vast amount of data that can be generated by these sensors. This is crucial because the system needs to generate and update its internal models in real-time. There should be little difference between the training and execution phases; the system should be continually learning, or engaged in "life-long learning". This paper explores research into incremental learning systems such as nearest neighbor, Bayesian classifiers, and fuzzy c-means clustering.

Teleoperation is the currently accepted method of control of military unmanned ground vehicles (UGVs) in the field. Degraded communications affects the operator's tasks of driving, navigating and maintaining UGV situation awareness. A potential approach to address this challenge is to provide the UGV with local autonomy to generate driving commands (translation and rotation rates). This paper describes an experiment and preliminary results comparing "point-and-go" supervisory control in which the operator designates a goal point on the 2D driving display to teleoperation as a function of communications degradation and terrain roughness. Three methods of visual supervisory control were tested (visual dead reckoning and two visual sevoing methods) and compared to teleoperation.

Terrains have widely varying visual appearance depending on the type of foliage, season, current weather conditions, recent precipitation, time of day, relative direction of lighting, presence of man-made structures and artifacts, landscaping, etc. It is difficult, if not impossible, to specify in advance the appearance of the different terrains that will be encountered while operating a robot in urban or rural environments. Yet people, having accumulated wide-ranging experience, have little trouble recognizing familiar terrain types and learning to recognize new, previously unfamiliar, terrains. Robots typically accumulate experience in "chunks" and do not have the luxury of years of training. This paper presents recent results in sequential learning methods applied to robot terrain recognition. In this paper we explore different sequential learning problem formulations and alternative machine learning algorithms. The investigations are based on the same data set. We report on the initial development of an incremental fuzzy c-means clustering algorithm capable of learning new information. We report on an approach to convert regression tree modeling, normally a batch learning method, to batch-incremental learning. We investigate issues in formulating the sequential learning problem and the performance of these algorithms. We also compare performance to four incremental learning classifiers. All investigations were conducted using the same set of image features, extracted from on-board video from a small robot traversing different terrains.

This paper presents an exploration of methods for estimating terrain trafficability from visual appearance. Two different sets of data are used. The first set is extracted from video sequences and has a small number of different terrains. A fuzzy c-means clustering algorithm is used to predict terrain type. The second set is derived from high-resolution still images and has a large variety of terrains. A decision tree algorithm is used to provide a subjective assessment of trafficability. A variety of local features are explored, based on color and texture, as input to the learning algorithms.

Engineered or hard-coded autonomous behaviors tend to be "brittle," working for a narrow range of conditions but failing outside that range. Trainable robots capable of learning and adapting to new environments and conditions have the potential for greater robustness and reusability. Trainable robots would not be restricted to learning from their own experience, but could potentially integrate models or lessons learned by other similar robots operating in different conditions, thus achieving a "learning force multiplier". In this research we began an investigation of issues and methods in robot learning, from definition of the learning objective, training methods, learning algorithms, and integration of models or lessons from multiple training sessions. Our objective in this initial research was not to develop new robot learning technologies, but to explore issues and approaches across all aspects of robot learning. In this stage of the project, we focused on learning to see, specifically learning to discriminate between "Go" and "NoGo" terrain.

This paper presents a method to forecast terrain trafficability from visual appearance. During training, the system identifies a set of image chips (or exemplars) that span the range of terrain appearance. Each chip is assigned a vector tag of vehicle-terrain interaction characteristics that are obtained from on-board sensors and simple performance models, as the vehicle traverses the terrain. The system uses the exemplars to segment images into regions, based on visual similarity to the terrain patches observed during training, and assigns the appropriate vehicle-terrain interaction tag to them. This methodology will therefore allow the online forecasting of vehicle performance on upcoming terrain. Currently, we are using fuzzy c-means clustering and exploring a number of different features for characterizing the visual appearance of the terrain.

This paper presents a method to forecast terrain trafficability from visual appearance. During training, the system identifies a set of image chips (or exemplars) that span the range of terrain appearance. Each chip is assigned a vector tag of vehicle-terrain interaction characteristics that are obtained from simple performance models and on-board sensors, as the vehicle traverses the terrain. The system uses the exemplars to segment images into regions, based on visual similarity to the terrain patches observed during training, and assigns the appropriate vehicle-terrain interaction tag to them. This methodology will therefore allow the online forecasting of vehicle performance on upcoming terrain. Currently, the system uses a fuzzy c-means clustering algorithm for training. In this paper, we explore a number of different features for characterizing the visual appearance of the terrain and measure their effect on the prediction of vehicle performance.

This paper presents an algorithm for online image-based terrain classification that mimics a human supervisor's segmentation and classification of training images into "Go" and "NoGo" regions. The algorithm identifies a set of image chips (or exemplars) in the training images that span the range of terrain appearance. It then uses the exemplars to segment novel images and assign a Go/NoGo classification. System parameters adapt to new inputs, providing a mechanism for learning. System performance is compared to that obtained via offline fuzzy c-means clustering and support vector machine classification.

: This paper presents a method to forecast terrain trafficability from visual appearance. During training, the system identifies a set of image chips (or exemplars) that span the range of terrain appearance and measures terrain trafficability characteristics as the vehicle traverses the terrain. Each chip is assigned a vector tag representing the measured vehicle-terrain interaction properties. After training, the system uses the exemplars to segment images into regions, based on visual similarity to terrain patches observed during training, and assigns the appropriate vehicle-terrain interaction tag to them. The system will therefore allow the online forecasting of vehicle performance on upcoming terrain.