Search

Autonomous navigation to a specified waypoint is traditionally accomplished with a layered stack of global path planning and local motion planning modules that generate feasible and obstacle-free trajectories. While these modules can be modified to meet task-specific constraints and user preferences, current modification procedures require substantial effort on the part of an expert roboticist with a great deal of technical training. In this paper, we simplify this process by inserting a Machine Learning module between the global path planning and local motion planning modules of an off-the shelf navigation stack. This model can be trained with human demonstrations of the preferred navigation behavior, using a training procedure based on Behavioral Cloning, allowing for an intuitive modification of the navigation policy by non-technical users to suit task-specific constraints. We find that our approach can successfully adapt a robot’s navigation behavior to become more like that of a demonstrator. Moreover, for a fixed amount of demonstration data, we find that the proposed technique compares favorably to recent baselines with respect to both navigation success rate and trajectory similarity to the demonstrator. Published version

This paper presents an approach and introduces new open-source tools that can be used to evaluate robotic mapping algorithms. Also described is an extensive subterranean mine rescue dataset based upon the DARPA Subterranean (SubT) challenge including professionally surveyed ground truth. Finally, some commonly available approaches are evaluated using this metric.

Robots are ideal surrogates for performing tasks that are dull, dirty, and dangerous. To fully achieve this ideal, a robotic teammate should be able to autonomously perform human-level tasks in unstructured environments where we do not want humans to go. In this paper, we take a step toward realizing that vision by introducing the integration of state of the art advancements in intelligence, perception, and manipulation on the RoMan (Robotic Manipulation) platform. RoMan is comprised of two 7 degree of freedom (DoF) limbs connected to a 1 DoF torso and mounted on a tracked base. Multiple lidars are used for navigation, and a stereo depth camera visualizes point clouds for grasping. Each limb has a 6 DoF force-torque sensor at the wrist, with a dexterous 3-finger gripper on one limb and a stronger 4-finger claw-like hand on the other. Tasks begin with an operator specifying a mission type, a desired final destination for the robot, and a general region where the robot should look for grasps. All other portions of the task are completed autonomously. This includes navigation, object identification and pose estimation (if the object is known) via deep learning or perception through search, fine maneuvering, grasp planning via grasp library, arm motion planning, and manipulation planning (e.g. dragging if the object is deemed too heavy to freely lift). Finally, we present initial test results on two notional tasks: clearing a road of debris such as a heavy tree or a pile of unknown light debris, and opening a hinged container to retrieve a bag inside it.

A primary goal for human-autonomy integration (HAI) is to balance the strengths of human and autonomy in order to achieve performance objectives more efficiently and robustly than either the human ...

We consider discriminative dictionary learning in a distributed online setting, where a team of networked robots aims to jointly learn both a common basis of the feature space and a classifier over this basis from sequentially observed signals. We formulate this problem as a distributed stochastic program with a non-convex objective and present a block variant of the Arrow-Hurwicz saddle point algorithm to solve it. Only neighboring nodes in the communications network need to exchange information, and we penalize the discrepency between the individual feature basis and classifiers using Lagrange multipliers. The application we consider is for a team of robots to collaboratively recognize objects of interest in dynamic environments. As a preliminary performance benchmark, we consider the problem of learning a texture classifier across a network of robots moving around an urban setting where separate training examples are sequentially observed at each robot. Results are shown for both a standard texture dataset and a new dataset from an urban training facility, and we compare the performance of the standard centralized construction to the new distributed algorithm for the case when distinct samples from all classes are seen by the robots. These experiments yield comparable performance between the decentralized and the centralized cases, demonstrating the proposed method's practical utility.

While deep reinforcement learning techniques have led to agents that are successfully able to learn to perform a number of tasks that had been previously unlearnable, these techniques are still susceptible to the longstanding problem of {\em reward sparsity}. This is especially true for tasks such as training an agent to play StarCraft II, a real-time strategy game where reward is only given at the end of a game which is usually very long. While this problem can be addressed through reward shaping, such approaches typically require a human expert with specialized knowledge. Inspired by the vision of enabling reward shaping through the more-accessible paradigm of natural-language narration, we investigate to what extent we can contextualize these narrations by grounding them to the goal-specific states. We present a mutual-embedding model using a multi-input deep-neural network that projects a sequence of natural language commands into the same high-dimensional representation space as corresponding goal states. We show that using this model we can learn an embedding space with separable and distinct clusters that accurately maps natural-language commands to corresponding game states . We also discuss how this model can allow for the use of narrations as a robust form of reward shaping to improve RL performance and efficiency., 10 pages, 3 figures. Published at SPIE 2019

Numerous, artificially intelligent, networked things will populate the battlefield of the future, operating in close collaboration with human warfighters, and fighting as teams in highly adversarial environments. This chapter explores the characteristics, capabilities, and intelligence required of such a network of intelligent things and humans—Internet of Battle Things (IoBT). The IoBT will experience unique challenges that are not yet well addressed by the current generation of artificial intelligence and machine learning.

The output encodings of neural nets determine the structure of the space in which inference occurs. Yet, they are generally given very little thought. It is common practice for neural nets to use 1-Hot encoding when training to discriminate among many classes. The primary exceptions to this are error correcting output codes, and semantic output encodings. Output encodings based upon semantic descriptors cause a net to learn responses for classes to which it has not been exposed, provided those classes may be characterized by the same semantic descriptors. This raises a number of questions, such as “Can a net implicitly learn encodings for unobserved classes in the absence of a semantic encoding, or any encoding that requires some form of prior knowledge and hand crafting?”. Also, are some output encodings better than others for learning these implicit encodings? In this paper, we will compare how effectively different non-semantic encodings are at causing a neural net to implicitly learn encodings for unobserved classes. Also, while evaluating the efficacy of these implicit encodings, we will look for evidence of a phenomenon akin to over-training. Specifically, as training on the observed classes occurs, we initially see improvement in how well the implicitly learned encodings can be used to differentiate among the classes which are unobserved during the net's training. However, as training continues to improve discrimination among the observed classes, the efficacy of the implicit codes either remains steady, or undergoes a degradation. This degradation is akin to the overtraining that one generally tries to guard against when training a neural net.

We consider policy evaluation in infinite-horizon discounted Markov decision problems with continuous compact state and action spaces. We reformulate this task as a compositional stochastic program with a function-valued decision variable that belongs to a reproducing kernel Hilbert space (RKHS). We approach this problem via a new functional generalization of stochastic quasi-gradient methods operating in tandem with stochastic sparse subspace projections. The result is an extension of gradient temporal difference learning that yields nonlinearly parameterized value function estimates of the solution to the Bellman evaluation equation. We call this method parsimonious kernel gradient temporal difference learning. Our main contribution is a memory-efficient nonparametric stochastic method guaranteed to converge exactly to the Bellman fixed point with probability 1 with attenuating step-sizes under the hypothesis that it belongs to the RKHS. Further, with constant step-sizes and compression budget, we establish mean convergence to a neighborhood and that the value function estimates have finite complexity. In the Mountain Car domain, we observe faster convergence to lower Bellman error solutions than existing approaches with a fraction of the required memory.

Understanding of intent is one of the most complex traits of highly efficient teams. Combining elements of verbal and non-verbal communication along with shared mental models about mission goals and team member capabilities, intent requires knowledge about both task and teammate. Beginning with the traditional models of communication, accounting for teaming factors, such as situation awareness, and incorporating the sensing, reasoning, and tactical capabilities available via autonomous systems, a revised model of team communication is needed to accurately describe the unique interactions and understanding of intent which will occur in human-robot teams. This paper focuses on examining the issue from a system capability viewpoint, identifying which system capabilities can mirror the abilities of humans through the sensor and computing strengths of autonomous systems, thus creating a team environment which is robust and adaptable while maintaining focus on mission goals.

In support of achieving better performance on autonomous mapping and exploration tasks by incorporating human input, we seek here to first characterize humans’ ability to recognize locations from limited visual information. Such a characterization is critical to the design of a human-in-the-loop system faced with deciding whether and when human input is useful. In this work, we develop a novel and practical place-recognition task that presents humans with video clips captured by a navigating ground robot. Using this task, we find experimentally that human performance does not seem to depend on factors such as clip length or familiarity with the scene and also that there is significant variability across subjects. Moreover, we find that humans significantly outperform a state-of-the-art computational solution to this problem, suggesting the utility of incorporating human input in autonomous mapping and exploration techniques.

Despite their attractiveness, popular perception is that techniques for nonparametric function approximation do not scale to streaming data due to an intractable growth in the amount of storage they require. To solve this problem in a memory-affordable way, we propose an online technique based on functional stochastic gradient descent in tandem with supervised sparsification based on greedy function subspace projections. The method, called parsimonious online learning with kernels (POLK), provides a controllable tradeoff? between its solution accuracy and the amount of memory it requires. We derive conditions under which the generated function sequence converges almost surely to the optimal function, and we establish that the memory requirement remains finite. We evaluate POLK for kernel multi-class logistic regression and kernel hinge-loss classification on three canonical data sets: a synthetic Gaussian mixture model, the MNIST hand-written digits, and the Brodatz texture database. On all three tasks, we observe a favorable tradeoff of objective function evaluation, classification performance, and complexity of the nonparametric regressor extracted the proposed method., Comment: Submitted to JMLR on 11/24/2016