Adversarial Imitation Learning Research Articles

BeTAIL: Behavior Transformer Adversarial Imitation Learning From Human Racing Gameplay

BeTAIL: Behavior Transformer Adversarial Imitation Learning From Human Racing Gameplay

Integrated reinforcement and imitation learning for tower crane lift path planning

Reinforcement learning (RL) has emerged as a promising solution method for crane-lift path planning. However, designing appropriate reward functions for tower crane (TC) operations remains particularly challenging. Poor design of reward functions can lead to non-executable lifting paths. This paper presents a framework combining imitation learning (IL) and RL to address the challenge. The framework comprises three steps: (1) designing a virtual environment consisting of construction site models and a TC model, (2) collecting expert demonstrations through virtual reality (VR) and pretraining through behavioral cloning (BC), and (3) refining the BC policies via generative adversarial imitation learning (GAIL) and proximal policy optimization (PPO). Using the paths generated by a PPO model as the baseline, the proposed BC + PPO + GAIL model exhibited better performance in both blind and nonblind lifting scenarios. This framework has been proven to generate realistic lifting paths mirroring crane operator behavior while ensuring efficiency and safety.

Modelling flight trajectories with multi-modal generative adversarial imitation learning

Modelling flight trajectories with multi-modal generative adversarial imitation learning

DeepCraft: imitation learning method in a cointelligent design to production process to deliver architectural scenarios

The recent advancements in digital technologies and artificial intelligence in the architecture, engineering, construction, and operation sector (AECO) have induced high demands on the digital skills of human experts, builders, and workers. At the same time, to satisfy the standards of the production-efficient AECO sector by reducing costs, energy, health risk, material resources, and labor demand through efficient production and construction methods such as design for manufacture and assembly (DfMA), it is necessary to resolve efficiency-related problems in mutual human‒machine collaborations. In this article, a method utilizing artificial intelligence (AI), namely, generative adversarial imitation learning (GAIL), is presented then evaluated in two independent experiments related to the processes of DfMA as an efficient human‒machine collaboration. These experiments include a) training the digital twin of a robot to execute a robotic toolpath according to human gestures and b) the generation of a spatial configuration driven by a human's design intent provided in a demonstration. The framework encompasses human intelligence and creativity, which the AI agent in the learning process observes, understands, learns, and imitates. For both experimental cases, the human demonstration, the agent's training, the toolpath execution, and the assembly configuration process are conducted digitally. Following the scenario generated by an AI agent in a digital space, physical assembly is undertaken by human builders as the next step. The implemented workflow successfully delivers the learned toolpath and scalable spatial assemblies, articulating human intelligence, intuition, and creativity in the cocreative design.

DiffAIL: Diffusion Adversarial Imitation Learning

Imitation learning aims to solve the problem of defining reward functions in real-world decision-making tasks. The current popular approach is the Adversarial Imitation Learning (AIL) framework, which matches expert state-action occupancy measures to obtain a surrogate reward for forward reinforcement learning. However, the traditional discriminator is a simple binary classifier and doesn't learn an accurate distribution, which may result in failing to identify expert-level state-action pairs induced by the policy interacting with the environment. To address this issue, we propose a method named diffusion adversarial imitation learning (DiffAIL), which introduces the diffusion model into the AIL framework. Specifically, DiffAIL models the state-action pairs as unconditional diffusion models and uses diffusion loss as part of the discriminator's learning objective, which enables the discriminator to capture better expert demonstrations and improve generalization. Experimentally, the results show that our method achieves state-of-the-art performance and significantly surpasses expert demonstration on two benchmark tasks, including the standard state-action setting and state-only settings.

Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective

Generative Adversarial Imitation Learning (GAIL) stands as a cornerstone approach in imitation learning. This paper investigates the gradient explosion in two types of GAIL: GAIL with deterministic policy (DE-GAIL) and GAIL with stochastic policy (ST-GAIL). We begin with the observation that the training can be highly unstable for DE-GAIL at the beginning of the training phase and end up divergence. Conversely, the ST-GAIL training trajectory remains consistent, reliably converging. To shed light on these disparities, we provide an explanation from a theoretical perspective. By establishing a probabilistic lower bound for GAIL, we demonstrate that gradient explosion is an inevitable outcome for DE-GAIL due to occasionally large expert-imitator policy disparity, whereas ST-GAIL does not have the issue with it. To substantiate our assertion, we illustrate how modifications in the reward function can mitigate the gradient explosion challenge. Finally, we propose CREDO, a simple yet effective strategy that clips the reward function during the training phase, allowing the GAIL to enjoy high data efficiency and stable trainability.

UAV Control Method Combining Reptile Meta-Reinforcement Learning and Generative Adversarial Imitation Learning

Reinforcement learning (RL) is pivotal in empowering Unmanned Aerial Vehicles (UAVs) to navigate and make decisions efficiently and intelligently within complex and dynamic surroundings. Despite its significance, RL is hampered by inherent limitations such as low sample efficiency, restricted generalization capabilities, and a heavy reliance on the intricacies of reward function design. These challenges often render single-method RL approaches inadequate, particularly in the context of UAV operations where high costs and safety risks in real-world applications cannot be overlooked. To address these issues, this paper introduces a novel RL framework that synergistically integrates meta-learning and imitation learning. By leveraging the Reptile algorithm from meta-learning and Generative Adversarial Imitation Learning (GAIL), coupled with state normalization techniques for processing state data, this framework significantly enhances the model’s adaptability. It achieves this by identifying and leveraging commonalities across various tasks, allowing for swift adaptation to new challenges without the need for complex reward function designs. To ascertain the efficacy of this integrated approach, we conducted simulation experiments within both two-dimensional environments. The empirical results clearly indicate that our GAIL-enhanced Reptile method surpasses conventional single-method RL algorithms in terms of training efficiency. This evidence underscores the potential of combining meta-learning and imitation learning to surmount the traditional barriers faced by reinforcement learning in UAV trajectory planning and decision-making processes.

BAGAIL: Multi-modal imitation learning from imbalanced demonstrations

Expert demonstrations in imitation learning often contain different behavioral modes, e.g., driving modes such as driving on the left, keeping the lane, and driving on the right in the driving tasks. Although most existing multi-modal imitation learning methods allow learning from demonstrations of multiple modes, they have strict constraints on the data of each mode, generally requiring a near data ratio of all modes. Otherwise, it tends to fall into a mode collapse or only learn the data distribution of the mode that has the largest data volume. To address the problem, an algorithm that balances real-fake loss and classification loss by modifying the output of the discriminator, referred to as BAlanced Generative Adversarial Imitation Learning (BAGAIL), is proposed. With this modification, the generator is only rewarded for generating real trajectories with correct modes. BAGAIL is therefore able to deal with imbalanced expert demonstrations and carry out efficient learning for each mode. The learning process of BAGAIL is divided into a pre-training stage and an imitation learning stage. During the pre-training stage, BAGAIL initializes the generator parameters by means of conditional Behavioral Cloning, laying the foundation for the direction of parameter optimization. During the imitation learning stage, BAGAIL optimizes the parameters by using the adversary between the generator and the modified discriminator so that the finally obtained policy can successfully learn the distribution of imbalanced expert data. The experiments showed that BAGAIL accurately distinguished different behavioral modes with imbalanced demonstrations. What is more, the learning result of each mode is close to the expert standard and more stable than other multi-modal imitation learning methods.

Imitation learning from imperfect demonstrations for AUV path tracking and obstacle avoidance

Autonomous underwater vehicle (AUV) is widely used for complex underwater tasks such as seafloor exploration. In recent years, deep reinforcement learning (DRL) has been introduced to the AUV control due to its capability to improve the autonomy of AUV. However, it is usually very difficult to design an effective reward function for the DRL methods. Generative adversarial imitation learning (GAIL) can allow AUVs to learn control policies from expert demonstrations instead of pre-defined reward functions, but suffers from the deficiency of requiring optimal expert demonstrations and not surpassing the provided demonstrations. This paper builds upon the GAIL algorithm for AUV learning control policies from expert demonstrations. We proposed an importance reweighting generative adversarial imitation learning (WGAIL) algorithm by using confidence scores to indicate the optimality of the demonstrated trajectories, which can facilitate AUVs to learn control policies from expert demonstrations of different levels. Our experimental results on a simulated AUV system modeling Sailfish 210 of our lab in the Gazebo simulation environment show that an AUV trained via WGAIL can achieve a better performance than the one trained via GAIL with different levels of expert sub-optimal demonstrations. Moreover, control policies trained via WGAIL in simple tasks can generalize better to complex tasks than those trained via GAIL, greatly extending the applicability of the AUV learning from expert demonstrations.

Generating Daily Activities with Need Dynamics

Daily activity data recording individuals’ various activities in daily life are widely used in many applications such as activity scheduling, activity recommendation, and policymaking. Though with high value, its accessibility is limited due to high collection costs and potential privacy issues. Therefore, simulating human activities to produce massive high-quality data is of great importance. However, existing solutions, including rule-based methods with simplified behavior assumptions and data-driven methods directly fitting real-world data, both cannot fully qualify for matching reality. In this article, motivated by the classic psychological theory, Maslow’s need theory describing human motivation, we propose a knowledge-driven simulation framework based on generative adversarial imitation learning. Our core idea is to model the evolution of human needs as the underlying mechanism that drives activity generation in the simulation model. Specifically, a hierarchical model structure that disentangles different need levels and the use of neural stochastic differential equations successfully capture the piecewise-continuous characteristics of need dynamics. Extensive experiments demonstrate that our framework outperforms the state-of-the-art baselines regarding data fidelity and utility. We also present the insightful interpretability of the need modeling. Moreover, privacy preservation evaluations validate that the generated data does not leak individual privacy. The code is available at https://github.com/tsinghua-fib-lab/Activity-Simulation-SAND .

GREEN PATH: an expert system for space planning and design by the generation of human trajectories

AbstractPublic space is usually conceived as where people live, perceive, and interact with other people. The environment affects people in several different ways as well. The impact of environmental problems on humans is significant, affecting all human activities, including health and socio-economic development. Thus, there is a need to rethink how space is used. Dealing with the important needs raised by climate emergency, pandemic and digitization, the contributions of this paper consist in the creation of opportunities for developing generative approaches to space design and utilization. It is proposed GREEN PATH, an intelligent expert system for space planning. GREEN PATH uses human trajectories and deep learning methods to analyse and understand human behaviour for offering insights to layout designers. In particular, a Generative Adversarial Imitation Learning (GAIL) framework hybridised with classical reinforcement learning methods is proposed. An example of the classical reinforcement learning method used is continuous penalties, which allow us to model the shape of the trajectories and insert a bias, which is necessary for the generation, into the training. The structure of the framework and the formalisation of the problem to be solved allow for the evaluation of the results in terms of generation and prediction. The use case is a chosen retail domain that will serve as a demonstrator for optimising the layout environment and improving the shopping experience. Experiments were assessed on shoppers’ trajectories obtained from four different stores, considering two years.

Real-Time On-the-Fly Motion Planning for Urban Air Mobility via Updating Tree Data of Sampling-Based Algorithms Using Neural Network Inference

In this study, we consider the problem of motion planning for urban air mobility applications to generate a minimal snap trajectory and trajectory that cost minimal time to reach a goal location in the presence of dynamic geo-fences and uncertainties in the urban airspace. We have developed two separate approaches for this problem because designing an algorithm individually for each objective yields better performance. The first approach that we propose is a decoupled method that includes designing a policy network based on a recurrent neural network for a reinforcement learning algorithm, and then combining an online trajectory generation algorithm to obtain the minimal snap trajectory for the vehicle. Additionally, in the second approach, we propose a coupled method using a generative adversarial imitation learning algorithm for training a recurrent-neural-network-based policy network and generating the time-optimized trajectory. The simulation results show that our approaches have a short computation time when compared to other algorithms with similar performance while guaranteeing sufficient exploration of the environment. In urban air mobility operations, our approaches are able to provide real-time on-the-fly motion re-planning for vehicles, and the re-planned trajectories maintain continuity for the executed trajectory. To the best of our knowledge, we propose one of the first approaches enabling one to perform an on-the-fly update of the final landing position and to optimize the path and trajectory in real-time while keeping explorations in the environment.

Self-Interested Coalitional Crowdsensing for Multi-Agent Interactive Environment Monitoring.

As a promising paradigm, mobile crowdsensing (MCS) takes advantage of sensing abilities and cooperates with multi-agent reinforcement learning technologies to provide services for users in large sensing areas, such as smart transportation, environment monitoring, etc. In most cases, strategy training for multi-agent reinforcement learning requires substantial interaction with the sensing environment, which results in unaffordable costs. Thus, environment reconstruction via extraction of the causal effect model from past data is an effective way to smoothly accomplish environment monitoring. However, the sensing environment is often so complex that the observable and unobservable data collected are sparse and heterogeneous, affecting the accuracy of the reconstruction. In this paper, we focus on developing a robust multi-agent environment monitoring framework, called self-interested coalitional crowdsensing for multi-agent interactive environment monitoring (SCC-MIE), including environment reconstruction and worker selection. In SCC-MIE, we start from a multi-agent generative adversarial imitation learning framework to introduce a new self-interested coalitional learning strategy, which forges cooperation between a reconstructor and a discriminator to learn the sensing environment together with the hidden confounder while providing interpretability on the results of environment monitoring. Based on this, we utilize the secretary problem to select suitable workers to collect data for accurate environment monitoring in a real-time manner. It is shown that SCC-MIE realizes a significant performance improvement in environment monitoring compared to the existing models.

The Application of Residual Connection-Based State Normalization Method in GAIL

In the domain of reinforcement learning (RL), deriving efficacious state representations and maintaining algorithmic stability are crucial for optimal agent performance. However, the inherent dynamism of state representations often complicates the normalization procedure. To overcome these challenges, we present an innovative RL framework that integrates state normalization techniques with residual connections and incorporates attention mechanisms into generative adversarial imitation learning (GAIL). This combination not only enhances the expressive capability of state representations, thereby improving the agent’s accuracy in state recognition, but also significantly mitigates the common issues of gradient dissipation and explosion. Compared to traditional RL algorithms, GAIL combined with the residual connection-based state normalization method empowers the agent to markedly reduce the exploration duration such that feedback concerning rewards in the current state can be provided in real time. Empirical evaluations demonstrate the superior efficacy of this methodology across various RL environments.

GAILPG: Multi-Agent Policy Gradient with Generative Adversarial Imitation Learning

GAILPG: Multi-Agent Policy Gradient with Generative Adversarial Imitation Learning

Imitation learning by state-only distribution matching

Imitation Learning from observation describes policy learning in a similar way to human learning. An agent’s policy is trained by observing an expert performing a task. Although many state-only imitation learning approaches are based on adversarial imitation learning, one main drawback is that adversarial training is often unstable and lacks a reliable convergence estimator. If the true environment reward is unknown and cannot be used to select the best-performing model, this can result in bad real-world policy performance. We propose a non-adversarial learning-from-observations approach, together with an interpretable convergence and performance metric. Our training objective minimizes the Kulback-Leibler divergence (KLD) between the policy and expert state transition trajectories which can be optimized in a non-adversarial fashion. Such methods demonstrate improved robustness when learned density models guide the optimization. We further improve the sample efficiency by rewriting the KLD minimization as the Soft Actor Critic objective based on a modified reward using additional density models that estimate the environment’s forward and backward dynamics. Finally, we evaluate the effectiveness of our approach on well-known continuous control environments and show state-of-the-art performance while having a reliable performance estimator compared to several recent learning-from-observation methods.

Robot Learning Incorporating Human Interventions in the Real World for Autonomous Surgical Endoscopic Camera Control

Recent studies in surgical robotics have focused on automating common surgical subtasks such as grasping and manipulation using deep reinforcement learning (DRL). In this work, we consider surgical endoscopic camera control for object tracking e.g. using the endoscopic camera manipulator (ECM) from the da Vinci Research Kit (dVRK) (Intuitive Inc., Sunnyvale, CA, USA) as a typical surgical robot learning task. A DRL policy for controlling the robot joint space movements is first trained in a simulation environment and then continues the learning in the real world. To speed up training and avoid significant failures (in this case, losing view of the object), human interventions are incorporated into the training process and regular DRL is combined with generative adversarial imitation learning (GAIL) to encourage imitating human behaviors. Experiments show that an average reward of 159.8 can be achieved within 1000 steps compared to only 121.8 without human interventions, and the view of the moving object is lost only twice during the training process out of 3 trials. These results show that human interventions can improve learning speed and significantly reduce failures during the training process.

A dynamic test scenario generation method for autonomous vehicles based on conditional generative adversarial imitation learning

Autonomous vehicles must be comprehensively evaluated before deployed in cities and highways. However, most existing evaluation approaches for autonomous vehicles are static and model environmental vehicles with predefined trajectories, which ignore the time-sequential interactions between the ego vehicle and environmental vehicles. In this paper, we propose a dynamic test scenario generation method to evaluate autonomous vehicles by modeling environmental vehicles as agents with human behavior and simulating the interaction process between the autonomous vehicle and environmental vehicles. Considering the multimodal features of traffic scenarios, we cluster the real-word traffic environments, and integrate the scenario class labels into the conditional generative adversarial imitation learning (CGAIL) model to generate different types of traffic scenarios. The proposed method is validated in a typical lane-change scenario that involves frequent interactions between ego vehicle and environmental vehicles. Results show that the proposed method further test autonomous vehicles’ ability to cope with dynamic scenarios, and can be used to infer the weaknesses of the tested vehicles.

Taking complementary advantages: Improving exploration via double self-imitation learning in procedurally-generated environments

Efficient exploration is the core issue of deep reinforcement learning. Although state-of-the-art exploration methods have achieved much progress in many tasks, they usually underperform in procedurally-generated environments, indicating the low capability of generalization of the agent. To address the problem, a self-imitation exploration approach for procedurally-generated environments, referred to as Double Self-Imitation Learning (DSIL), is proposed. DSIL screens out good history experiences of exploration by utilizing an episode scoring rule that considers local scores, global scores and external rewards. Then DSIL employs a cooperation strategy to reproduce the agent’s past good exploration behaviors by combining generative adversarial imitation learning (GAIL) and behavioral cloning (BC). Specifically, DSIL is composed of a reinforcement learning module and a discriminator. The discriminator generates intrinsic rewards by judging the similarity of the current state–action pairs to the past good exploration experiences. The policy of agent is optimized alternately by the BC task and the reinforcement learning algorithm in the GAIL task; meanwhile, the reinforcement learning module and the discriminator are updated alternately in the GAIL task. Experiments on several procedurally-generated environments demonstrated that the proposed DSIL significantly outperformed existing exploration approaches in sample efficiency and performance, that is, DSIL makes the agent have stronger generalization.

Sim-GAIL: A generative adversarial imitation learning approach of student modelling for intelligent tutoring systems

The continuous application of artificial intelligence (AI) technologies in online education has led to significant progress, especially in the field of Intelligent Tutoring Systems (ITS), online courses and learning management systems (LMS). An important research direction of the field is to provide students with customised learning trajectories via student modelling. Previous studies have shown that customisation of learning trajectories could effectively improve students’ learning experiences and outcomes. However, training an ITS that can customise students’ learning trajectories suffers from cold-start, time-consumption, human labour-intensity, and cost problems. One feasible approach is to simulate real students’ behaviour trajectories through algorithms, to generate data that could be used to train the ITS. Nonetheless, implementing high-accuracy student modelling methods that effectively address these issues remains an ongoing challenge. Traditional simulation methods, in particular, encounter difficulties in ensuring the quality and diversity of the generated data, thereby limiting their capacity to provide intelligent tutoring systems (ITS) with high-fidelity and diverse training data. We thus propose Sim-GAIL, a novel student modelling method based on generative adversarial imitation learning (GAIL). To the best of our knowledge, it is the first method using GAIL to address the challenge of lacking training data, resulting from the issues mentioned above. We analyse and compare the performance of Sim-GAIL with two traditional Reinforcement Learning-based and Imitation Learning-based methods using action distribution evaluation, cumulative reward evaluation, and offline-policy evaluation. The experiments demonstrate that our method outperforms traditional ones on most metrics. Moreover, we apply our method to a domain plagued by the cold-start problem, knowledge tracing (KT), and the results show that our novel method could effectively improve the KT model’s prediction accuracy in a cold-start scenario.