Many reinforcement learning algorithms use value functions to guide the search for better policies. These methods estimate the value of a single policy while generalizing across many states. The core idea of this paper is to flip this convention and estimate the value of many policies, for a single set of states. This approach opens up the possibility of performing direct gradient ascent in policy space without seeing any new data. The main challenge for this approach is finding a way to represent complex policies that facilitates learning and generalization. To address this problem, we introduce a scalable, differentiable fingerprinting mechanism that retains essential policy information in a concise embedding. Our empirical results demonstrate that combining these three elements (learned Policy Evaluation Network, policy fingerprints, gradient ascent) can produce policies that outperform those that generated the training data, in zero-shot manner.

One of the key reasons for the high sample complexity in reinforcement learning (RL) is the inability to transfer knowledge from one task to another. In standard multi-task RL settings, low-reward data collected while trying to solve one task provides little to no signal for solving that particular task and is hence effectively wasted. However, we argue that this data, which is uninformative for one task, is likely a rich source of information for other tasks. To leverage this insight and efficiently reuse data, we present Generalized Hindsight: an approximate inverse reinforcement learning technique for relabeling behaviors with the right tasks. Intuitively, given a behavior generated under one task, Generalized Hindsight returns a different task that the behavior is better suited for. Then, the behavior is relabeled with this new task before being used by an off-policy RL optimizer. Compared to standard relabeling techniques, Generalized Hindsight provides a substantially more efficient reuse of samples, which we empirically demonstrate on a suite of multi-task navigation and manipulation tasks. Videos and code can be accessed here: https://sites.google.com/view/generalized-hindsight.

Compared to rigid robots that are often studied in reinforcement learning, the physical characteristics of some sophisticated robots such as software or continuum are more complicated. Moreover, recent reinforcement learning methods are data-inefficient and can not be directly deployed to the robot without simulation. In this paper, we propose an efficient reinforcement learning method based on inexplicit prior knowledge in response to such problems. The method is firstly corroborated by simulation and employed directly in the real world. By using our method, we can achieve visual active tracking and distance maintenance of a tendon-driven robot which will be critical in minimally-invasive procedures.

Driver distraction a significant risk to driving safety. Apart from spatial domain, research on temporal inattention is also necessary. In this paper, we propose an actor-critic method - Attention-based Twin Delayed Deep Deterministic policy gradient (ATD3) algorithm to approximate a driver's action according to observations and measure the driver's attention allocation for consecutive time steps in car-following model. Considering reaction time, we construct the attention mechanism in the actor network to capture temporal dependencies of consecutive observations. In the critic network, we employ Twin Delayed Deep Deterministic policy gradient algorithm (TD3) to address overestimated value estimates persisting in the actor-critic algorithm. We conduct experiments on real-world vehicle trajectory datasets and show that the accuracy of our proposed approach outperforms seven baseline algorithms. Moreover, the results reveal that the attention of the drivers in smooth vehicles is uniformly distributed in previous observations while they keep their attention to recent observations when sudden decreases of relative speeds occur. This study is the first contribution to drivers' temporal attention.

Visual Indoor Navigation (VIN) task has drawn increasing attentions from the data-driven machine learning communities especially with the recent reported success from learning-based methods. Due to the innate complexity of this task, researchers have tried approaching the problem from a variety of different angles, the full scope of which has not yet been captured within an overarching report. In this survey, we discuss the representative work of learning-based approaches for visual navigation and its related tasks. Firstly, we summarize the current work in terms of task representations and applied methods along with their properties. We then further identify and discuss lingering issues impeding the performance of VIN tasks and motivate future research in these key areas worth exploring in the future for the community.

Mean field theory provides an effective way of scaling multiagent reinforcement learning algorithms to environments with many agents that can be abstracted by a virtual mean agent. In this paper, we extend mean field multiagent algorithms to multiple types. The types enable the relaxation of a core assumption in mean field games, which is that all agents in the environment are playing almost similar strategies and have the same goal. We conduct experiments on three different testbeds for the field of many agent reinforcement learning, based on the standard MAgents framework. We consider two different kinds of mean field games: a) Games where agents belong to predefined types that are known a priori and b) Games where the type of each agent is unknown and therefore must be learned based on observations. We introduce new algorithms for each type of game and demonstrate their superior performance over state of the art algorithms that assume that all agents belong to the same type and other baseline algorithms in the MAgent framework.

This paper considers if a reward function learned via inverse reinforcement from a human expert can be used as a feedback intervention to alter future human performance as desired (i.e., human to human apprenticeship learning). To learn reward functions two new algorithms are developed: a kernel-based inverse reinforcement learning algorithm and a Monte Carlo reinforcement learning algorithm. The algorithms are benchmarked against well-known alternatives within their respective corpus and are shown to outperform in terms of efficiency and optimality. To test the feedback intervention two randomized experiments are performed with 3,256 human participants. The experimental results demonstrate with significance that the rewards learned from "expert" individuals are effective as feedback interventions. In addition to the algorithmic contributions and successful experiments, the paper also describes three reward function modifications to improve reward function feedback interventions for humans.

While robot learning has demonstrated promising results for enabling robots to automatically acquire new skills, a critical challenge in deploying learning-based systems is scale: acquiring enough data for the robot to effectively generalize broadly. Imitation learning, in particular, has remained a stable and powerful approach for robot learning, but critically relies on expert operators for data collection. In this work, we target this challenge, aiming to build an imitation learning system that can continuously improve through autonomous data collection, while simultaneously avoiding the explicit use of reinforcement learning, to maintain the stability, simplicity, and scalability of supervised imitation. To accomplish this, we cast the problem of imitation with autonomous improvement into a multi-task setting. We utilize the insight that, in a multi-task setting, a failed attempt at one task might represent a successful attempt at another task. This allows us to leverage the robot's own trials as demonstrations for tasks other than the one that the robot actually attempted. Using an initial dataset of multi-task demonstration data, the robot autonomously collects trials which are only sparsely labeled with a binary indication of whether the trial accomplished any useful task or not. We then embed the trials into a learned latent space of tasks, trained using only the initial demonstration dataset, to draw similarities between various trials, enabling the robot to achieve one-shot generalization to new tasks. In contrast to prior imitation learning approaches, our method can autonomously collect data with sparse supervision for continuous improvement, and in contrast to reinforcement learning algorithms, our method can effectively improve from sparse, task-agnostic reward signals.

The development of autonomous robotic systems that can learn from human demonstrations to imitate a desired behavior - rather than being manually programmed - has huge technological potential. One major challenge in imitation learning is the correspondence problem: how to establish corresponding states and actions between expert and learner, when the embodiments of the agents are different (morphology, dynamics, degrees of freedom, etc.). Many existing approaches in imitation learning circumvent the correspondence problem, for example, kinesthetic teaching or teleoperation, which are performed on the robot. In this work we explicitly address the correspondence problem by introducing a distance measure between dissimilar embodiments. This measure is then used as a loss function for static pose imitation and as a feedback signal within a model-free deep reinforcement learning framework for dynamic movement imitation between two anthropomorphic robotic arms in simulation. We find that the measure is well suited for describing the similarity between embodiments and for learning imitation policies by distance minimization.

Partial monitoring is a rich framework for sequential decision making under uncertainty that generalizes many well known bandit models, including linear, combinatorial and dueling bandits. We introduce information directed sampling (IDS) for stochastic partial monitoring with a linear reward and observation structure. IDS achieves adaptive worst-case regret rates that depend on precise observability conditions of the game. Moreover, we prove lower bounds that classify the minimax regret of all finite games into four possible regimes. IDS achieves the optimal rate in all cases up to logarithmic factors, without tuning any hyper-parameters. We further extend our results to the contextual and the kernelized setting, which significantly increases the range of possible applications.