Guided exploration with expert demonstrations improves data efficiency for reinforcement learning, but current algorithms often overuse expert information. We propose a novel algorithm to speed up Q-learning with the help of a limited amount of imperfect expert demonstrations. The algorithm avoids excessive reliance on expert data by relaxing the optimal expert assumption and gradually reducing the usage of uninformative expert data. Experimentally, we evaluate our approach on a sparse-reward chain environment and six more complicated Atari games with delayed rewards. With the proposed methods, we can achieve better results than Deep Q-learning from Demonstrations (Hester et al., 2017) in most environments.

Actor-critic (AC) algorithms are a class of model-free deep reinforcement learning algorithms, which have proven their efficacy in diverse domains, especially in solving continuous control problems. Improvement of exploration (action entropy) and exploitation (expected return) using more efficient samples is a critical issue in AC algorithms. A basic strategy of a learning algorithm is to facilitate indiscriminately exploring all of the environment state space, as well as to encourage exploring rarely visited states rather than frequently visited one. Under this strategy, we propose a new method to boost exploration through an intrinsic reward, based on measurement of a state's novelty and the associated benefit of exploring the state (with regards to policy optimization), altogether called plausible novelty. With incentivized exploration of plausible novel states, an AC algorithm is able to improve its sample efficiency and hence training performance. The new method is verified by extensive simulations of continuous control tasks of MuJoCo environments on a variety of prominent off-policy AC algorithms.

Bus bunching is a natural-occurring phenomenon that undermines the efficiency and stability of the public transportation system. The mainstream solutions control the bus to intentionally stay longer at certain stations. Existing control methods include conventional methods that provide a formula to calculate the control time and reinforcement learning (RL) methods that determine the control policy through repeated interactions with the system. In this paper, we propose an integrated proximal policy optimization model with dual-headway (IPPO-DH). IPPO-DH integrates the conventional headway control with reinforcement learning, so that it acquires the advantages of both algorithms -- it is more efficient in normal environments and more stable in harsh ones. To demonstrate such an advantage, we design a bus simulation environment and compare IPPO-DH with RL and several conventional methods. The results show that the proposed model maintains the application value of the conventional method by avoiding the instability of the RL method in certain environments, and improves the efficiency compared with the conventional control, shedding new light on real-world bus transit system optimization.

Humans are capable of abstracting various tasks as different combinations of multiple attributes. This perspective of compositionality is vital for human rapid learning and adaption since previous experiences from related tasks can be combined to generalize across novel compositional settings. In this work, we aim to achieve zero-shot policy generalization of Reinforcement Learning (RL) agents by leveraging the task compositionality. Our proposed method is a meta- RL algorithm with disentangled task representation, explicitly encoding different aspects of the tasks. Policy generalization is then performed by inferring unseen compositional task representations via the obtained disentanglement without extra exploration. The evaluation is conducted on three simulated tasks and a challenging real-world robotic insertion task. Experimental results demonstrate that our proposed method achieves policy generalization to unseen compositional tasks in a zero-shot manner.

Powered by acoustics, existing therapeutic and diagnostic procedures will become less invasive and new methods will become available that have never been available before. Acoustically driven microrobot navigation based on microbubbles is a promising approach for targeted drug delivery. Previous studies have used acoustic techniques to manipulate microbubbles in vitro and in vivo for the delivery of drugs using minimally invasive procedures. Even though many advanced capabilities and sophisticated control have been achieved for acoustically powered microrobots, there remain many challenges that remain to be solved. In order to develop the next generation of intelligent micro/nanorobots, it is highly desirable to conduct accurate identification of the micro-nanorobots and to control their dynamic motion autonomously. Here we use reinforcement learning control strategies to learn the microrobot dynamics and manipulate them through acoustic forces. The result demonstrated for the first time autonomous acoustic navigation of microbubbles in a microfluidic environment. Taking advantage of the benefit of the second radiation force, microbubbles swarm to form a large swarm, which is then driven along the desired trajectory. More than 100 thousand images were used for the training to study the unexpected dynamics of microbubbles. As a result of this work, the microrobots are validated to be controlled, illustrating a good level of robustness and providing computational intelligence to the microrobots, which enables them to navigate independently in an unstructured environment without requiring outside assistance.

This paper analyzes the simulation to reality gap in reinforcement learning (RL) cyber-physical systems with fractional delays (i.e. delays that are non-integer multiple of the sampling period). The consideration of fractional delay has important implications on the nature of the cyber-physical system considered. Systems with delays are non-Markovian, and the system state vector needs to be extended to make the system Markovian. We show that this is not possible when the delay is in the output, and the problem would always be non-Markovian. Based on this analysis, a sampling scheme is proposed that results in efficient RL training and agents that perform well in realistic multirotor unmanned aerial vehicle simulations. We demonstrate that the resultant agents do not produce excessive oscillations, which is not the case with RL agents that do not consider time delay in the model.

While Unmanned Aerial Vehicles (UAVs) are increasingly deployed in several missions, their inability of reliable and consistent autonomous landing poses a major setback for deploying such systems truly autonomously. In this paper we present an autonomous UAV landing system for landing on a moving platform. In contrast to existing attempts, the proposed system relies only on the camera sensor, and has been designed as lightweight as possible. The proposed system can be deployed on a low power platform as part of the drone payload, whilst being indifferent to any external communication or any other sensors. The system relies on a Neural Network (NN) based controller, for which a target and environment agnostic simulator was created, used in training and testing of the proposed system, via Reinforcement Learning (RL) and Proximal Policy optimization (PPO) to optimally control and steer the drone towards landing on the target. Through real-world testing, the system was evaluated with an average deviation of 15cm from the center of the target, for 40 landing attempts.

Recent progress in state-only imitation learning extends the scope of applicability of imitation learning to real-world settings by relieving the need for observing expert actions. However, existing solutions only learn to extract a state-to-action mapping policy from the data, without considering how the expert plans to the target. This hinders the ability to leverage demonstrations and limits the flexibility of the policy. In this paper, we introduce Decoupled Policy Optimization (DePO), which explicitly decouples the policy as a high-level state planner and an inverse dynamics model. With embedded decoupled policy gradient and generative adversarial training, DePO enables knowledge transfer to different action spaces or state transition dynamics, and can generalize the planner to out-of-demonstration state regions. Our in-depth experimental analysis shows the effectiveness of DePO on learning a generalized target state planner while achieving the best imitation performance. We demonstrate the appealing usage of DePO for transferring across different tasks by pre-training, and the potential for co-training agents with various skills.

Batch reinforcement learning (BRL) is an emerging research area in the RL community. It learns exclusively from static datasets (i.e. replay buffers) without interaction with the environment. In the offline settings, existing replay experiences are used as prior knowledge for BRL models to find the optimal policy. Thus, generating replay buffers is crucial for BRL model benchmark. In our B2RL (Building Batch RL) dataset, we collected real-world data from our building management systems, as well as buffers generated by several behavioral policies in simulation environments. We believe it could help building experts on BRL research. To the best of our knowledge, we are the first to open-source building datasets for the purpose of BRL learning.

Offline reinforcement learning (Offline RL) suffers from the innate distributional shift as it cannot interact with the physical environment during training. To alleviate such limitation, state-based offline RL leverages a learned dynamics model from the logged experience and augments the predicted state transition to extend the data distribution. For exploiting such benefit also on the image-based RL, we firstly propose a generative model, S2P (State2Pixel), which synthesizes the raw pixel of the agent from its corresponding state. It enables bridging the gap between the state and the image domain in RL algorithms, and virtually exploring unseen image distribution via model-based transition in the state space. Through experiments, we confirm that our S2P-based image synthesis not only improves the image-based offline RL performance but also shows powerful generalization capability on unseen tasks.