We propose DeepExplorer, a simple and lightweight metric-free exploration method for topological mapping of unknown environments. It performs task and motion planning (TAMP) entirely in image feature space. The task planner is a recurrent network using the latest image observation sequence to hallucinate a feature as the next-best exploration goal. The motion planner then utilizes the current and the hallucinated features to generate an action taking the agent towards that goal. The two planners are jointly trained via deeply-supervised imitation learning from expert demonstrations. During exploration, we iteratively call the two planners to predict the next action, and the topological map is built by constantly appending the latest image observation and action to the map and using visual place recognition (VPR) for loop closing. The resulting topological map efficiently represents an environment's connectivity and traversability, so it can be used for tasks such as visual navigation. We show DeepExplorer's exploration efficiency and strong sim2sim generalization capability on large-scale simulation datasets like Gibson and MP3D. Its effectiveness is further validated via the image-goal navigation performance on the resulting topological map. We further show its strong zero-shot sim2real generalization capability in real-world experiments. The source code is available at \url{https://ai4ce.github.io/DeepExplorer/}.

In this paper, we formulate a Collaborative Pure Exploration in Kernel Bandit problem (CoPE-KB), which provides a novel model for multi-agent multi-task decision making under limited communication and general reward functions, and is applicable to many online learning tasks, e.g., recommendation systems and network scheduling. We consider two settings of CoPE-KB, i.e., Fixed-Confidence (FC) and Fixed-Budget (FB), and design two optimal algorithms CoopKernelFC (for FC) and CoopKernelFB (for FB). Our algorithms are equipped with innovative and efficient kernelized estimators to simultaneously achieve computation and communication efficiency. Matching upper and lower bounds under both the statistical and communication metrics are established to demonstrate the optimality of our algorithms. The theoretical bounds successfully quantify the influences of task similarities on learning acceleration and only depend on the effective dimension of the kernelized feature space. Our analytical techniques, including data dimension decomposition, linear structured instance transformation and (communication) round-speedup induction, are novel and applicable to other bandit problems. Empirical evaluations are provided to validate our theoretical results and demonstrate the performance superiority of our algorithms.

We present a lightweight, decentralized algorithm for navigating multiple nonholonomic agents through challenging environments with narrow passages. Our key idea is to allow agents to yield to each other in large open areas instead of narrow passages, to increase the success rate of conventional decentralized algorithms. At pre-processing time, our method computes a medial axis for the freespace. A reference trajectory is then computed and projected onto the medial axis for each agent. During run time, when an agent senses other agents moving in the opposite direction, our algorithm uses the medial axis to estimate a Point of Impact (POI) as well as the available area around the POI. If the area around the POI is not large enough for yielding behaviors to be successful, we shift the POI to nearby large areas by modulating the agent's reference trajectory and traveling speed. We evaluate our method on a row of 4 environments with up to 15 robots, and we find our method incurs a marginal computational overhead of 10-30 ms on average, achieving real-time performance. Afterward, our planned reference trajectories can be tracked using local navigation algorithms to achieve up to a $100\%$ higher success rate over local navigation algorithms alone.

Multi-agent reinforcement learning (MARL) is the study of the learning dynamics of strategic agents that coexist in a shared environment, and is one of the important frontiers of machine learning and control. In this paper, we study MARL in stochastic games, also known as Markov games, a multi-agent generalization of Markov decision problems (MDPs) in which the cost-relevant history of the system is summarized by a state variable Shapley [1953]. Due to its ability to model both dynamic inter-temporal choice as well as strategic interaction, the stochastic games model has long been a popular framework for studying multi-agent learning Littman [1994]. In comparison to single-agent reinforcement learning, analysis of MARL is difficult due to several challenges inherent to multi-agent systems, including non-stationarity, conflicting interests, and decentralized information. As a result, fundamental understanding of multi-agent reinforcement learning theory has lagged behind its single-agent counterpart Zhang et al. [2021].

The field of research aims to create agents that use visual sensors for solving complex tasks or aid humans by learning to perceive, communicate, and act in their environment. Humans in the loop make the goal very difficult since the dynamics of the environment are changeable, and human interactions can lead to unexpected events. Towards better collaboration between agents and humans, agents must be able to perform localization of any point in space that reflects the characteristics of human's perception of 3D space Cirik et al. [2020]. Since the visual sensors for the agents are commonly RGB sensors employed with partial Field-of-View (FoV), we would need to train these agents to perceive how humans see from these views. Communication with these agents will almost always necessitate the agents to navigate to view a common referential FoV in the scene so that the human can instruct the agents with the shared contexts. Challenge arises since the point of interest could be any point in the scene, and many points in the scene will not correspond to easily named objects. So far, many embodied agents being researched use either partial FoVs or directly use panoramic images that are hard for human observers to understand. We believe that embodied agents should be able to look around and localize in various views that human observers might be looking at. We approach this problem by introducing a new task, namely the FindView task, to evaluate and benchmark the agents (Figure 1).

Value-decomposition methods, which reduce the difficulty of a multi-agent system by decomposing the joint state-action space into local observation-action spaces, have become popular in cooperative multi-agent reinforcement learning (MARL). However, value-decomposition methods still have the problems of tremendous sample consumption for training and lack of active exploration. In this paper, we propose a scalable value-decomposition exploration (SVDE) method, which includes a scalable training mechanism, intrinsic reward design, and explorative experience replay. The scalable training mechanism asynchronously decouples strategy learning with environmental interaction, so as to accelerate sample generation in a MapReduce manner. For the problem of lack of exploration, an intrinsic reward design and explorative experience replay are proposed, so as to enhance exploration to produce diverse samples and filter non-novel samples, respectively. Empirically, our method achieves the best performance on almost all maps compared to other popular algorithms in a set of StarCraft II micromanagement games. A data-efficiency experiment also shows the acceleration of SVDE for sample collection and policy convergence, and we demonstrate the effectiveness of factors in SVDE through a set of ablation experiments.

This article presents the design and the implementation of a cloud system for knowledge-based autonomous interaction devised for Social Robots and other conversational agents. The system is particularly convenient for low-cost robots and devices: it can be used as a stand-alone dialogue system or as an integration to provide "background" dialogue capabilities to any preexisting Natural Language Processing ability that the robot may already have as part of its basic skills. By connecting to the cloud, developers are provided with a sustainable solution to manage verbal interaction through a network connection, with about 3,000 topics of conversation ready for "chit-chatting" and a library of pre-cooked plans that only needs to be grounded into the robot's physical capabilities. The system is structured as a set of REST API endpoints so that it can be easily expanded by adding new APIs to improve the capabilities of the clients connected to the cloud. Another key feature of the system is that it has been designed to make the development of its clients straightforward: in this way, multiple robots and devices can be easily endowed with the capability of autonomously interacting with the user, understanding when to perform specific actions, and exploiting all the information provided by cloud services. The article outlines and discusses the results of the experiments performed to assess the system's performance in terms of response time, paving the way for its use both for research and market solutions. Links to repositories with clients for ROS and popular robots such as Pepper and NAO are available on request.

This paper reports on the second GENEA Challenge to benchmark data-driven automatic co-speech gesture generation. Participating teams used the same speech and motion dataset to build gesture-generation systems. Motion generated by all these systems was rendered to video using a standardised visualisation pipeline and evaluated in several large, crowdsourced user studies. Unlike when comparing different research papers, differences in results are here only due to differences between methods, enabling direct comparison between systems. The dataset was based on 18 hours of full-body motion capture, including fingers, of different persons engaging in a dyadic conversation. Ten teams participated in the challenge across two tiers: full-body and upper-body gesticulation. For each tier, we evaluated both the human-likeness of the gesture motion and its appropriateness for the specific speech signal. Our evaluations decouple human-likeness from gesture appropriateness, which has been a difficult problem in the field. The evaluation results are a revolution, and a revelation. Some synthetic conditions are rated as significantly more human-like than human motion capture. To the best of our knowledge, this has never been shown before on a high-fidelity avatar. On the other hand, all synthetic motion is found to be vastly less appropriate for the speech than the original motion-capture recordings. We also find that conventional objective metrics do not correlate well with subjective human-likeness ratings in this large evaluation. The one exception is the Fr\'echet gesture distance (FGD), which achieves a Kendall's tau rank correlation of around -0.5. Based on the challenge results we formulate numerous recommendations for system building and evaluation.

Natural behavior consists of dynamics that are complex and unpredictable, especially when trying to predict many steps into the future. While some success has been found in building representations of behavior under constrained or simplified task-based conditions, many of these models cannot be applied to free and naturalistic settings where behavior becomes increasingly hard to model. In this work, we develop a multi-task representation learning model for behavior that combines two novel components: (i) an action-prediction objective that aims to predict the distribution of actions over future timesteps, and (ii) a multi-scale architecture that builds separate latent spaces to accommodate short-and long-term dynamics. After demonstrating the ability of the method to build representations of both local and global dynamics in realistic robots in varying environments and terrains, we apply our method to the MABe 2022 Multi-agent behavior challenge, where our model ranks first overall (top rank over all 13 tasks), first on all sequence-level tasks, and 1st or 2nd on 7 out of 9 frame-level tasks. In all of these cases, we show that our model can build representations that capture the many different factors that drive behavior and solve a wide range of downstream tasks.

Reinforcement learning (RL) has become widely adopted in robot control. Despite many successes, one major persisting problem can be very low data efficiency. One solution is interactive feedback, which has been shown to speed up RL considerably. As a result, there is an abundance of different strategies, which are, however, primarily tested on discrete grid-world and small scale optimal control scenarios. In the literature, there is no consensus about which feedback frequency is optimal or at which time the feedback is most beneficial. To resolve these discrepancies we isolate and quantify the effect of feedback frequency in robotic tasks with continuous state and action spaces. The experiments encompass inverse kinematics learning for robotic manipulator arms of different complexity. We show that seemingly contradictory reported phenomena occur at different complexity levels. Furthermore, our results suggest that no single ideal feedback frequency exists. Rather that feedback frequency should be changed as the agent's proficiency in the task increases.