Recent advancements in multi-robot navigation have explored methods that combine Large Language Models (LLMs) for tasks like scene understanding or high-level decision-making. However, these approaches face challenges with high inference latency and potential hallucinations. To address these challenges, we propose a knowledge-driven Reinforcement Learning (RL) framework, GUIDER, that utilizes an LLM in two different offline roles. First, we leverage the LLM as an offline knowledge source. Its expertise is distilled into a compact model, which is applied only when the RL agent is uncertain about its own value estimates and the model itself is confident in its prediction. Additionally, we utilize the LLM as an offline semantic engine. This process translates the LLM's high-level understanding of situational risk into a dynamic adjustment of the RL agent's behavioral style, evolving a function that optimally balances conservative and aggressive actions. We conduct extensive experiments in both terrestrial and maritime settings. Across all maritime scenarios (3–12 robots), GUIDER improves the task success rate and reduces the collision rate significantly compared to the state-of-the-art RL-based multi-robot navigation methods.

Due to recent advances in large language models and robotics, social robots will potentially play an important role in people’s daily lives soon, and are expected to improve dynamic multi-party group discussions in social scenarios. In this paper, we developed a system to assist dynamic group discussion with our social robot Haru. Our system is composed of three modules: a Dialogue Assistance module via integrating Haru with large language models which facilitates Haru to be an embodied chatbot; a Balancing and Welcoming Behavior module to improve users’ engagement and welcome new users to join the discussion with verbal behaviors; an Autonomous Eye Gazing module to show politeness during group discussion, e.g., gazing to the talking user or the less-engaging user to encourage her, looking to the new comer when she joins the discussion, gazing via eyeball movement when the current speaking user is close to the previous one. The autonomous eye gazing behavior was first trained via deep reinforcement learning in simulation and transferred to physical Haru in the real world. Results of our user study with 50 subjects show the significant performance of our system in assisting dynamic group discussion.

Socially assistive robots are potentially to be integrated with human daily lives in the near future, and expected to be able to improve group dynamics when interacting with groups of people in social settings. In this paper, we developed a system with desktop robot Haru to assist group discussions. The system consists of three modules: a dialogue assistance module which facilitates Haru to speak to users and answer questions in a free way; a dialogue balance module to encourage participation of users in the discussion with verbal behaviors; an autonomous gazing behavior module trained via deep reinforcement learning in simulation and deployed on physical Haru in reality, which can show politeness during group discussion, e.g., gazing to the speaking member, looking to the middle when both members are talking or silent, looking at the least spoken person when encouraging her. Results of user study with 40 subjects show the significant effectiveness of our system in assisting group discussion.

In this paper, building on recent advances in the fields of gaming AI and social robotics, we present a new approach to facilitate the social robot Haru to imitate game strategies from human players’ demonstrated trajectories and evaluative feedback in a real-time two-player game. Our research shows that Haru is able to learn and imitate human different game strategies from human players in a human time scale. In addition, our results show that human evaluative feedback plays an important role in allowing Haru to obtain a better performance via our method than human player’s demonstrations. Finally, results of our user study indicate that Haru imitating human player’s game strategies via our method is perceived to be more human-like and have better game performance and experience than self-learning from pre-defined reward functions via traditional deep reinforcement learning.

Social robots are gradually integrating into human’s daily lives. Storytelling by social robots could bring a different experience to users through non-verbal and emotional capabilities compared to text-only one. However, as user needs and preferences over storytelling might change over time during long-term interaction with social robots, it is important for social robots to learn from social interactions with human users in real-time. In this paper, we propose to allow our social robot Haru to learn personalized storytelling styles for different human user’s emotional states via human-centered reinforcement learning using the reward provided and delivered by directly interaction with the user explicitly. Results of our user study show that Haru can learn to adapt its storytelling style for detected human emotional states in a few number of interactions, and was perceived to have a better storytelling performance, experience and impact than a neutral one.

Generative adversarial imitation learning (GAIL) — a general model-free imitation learning method, allows robots to directly learn policies from expert trajectories in large environments. However, GAIL shares the limitation of other imitation learning methods that they can seldom surpass the performance of demonstrations. In this paper, to address the limit of GAIL, we propose GAN-based interactive reinforcement learning (GAIRL) from demonstrations and human evaluative feedback, by combining the advantages of GAIL and interactive reinforcement learning. We test GAIRL in six physics-based control tasks, ranging from simple low-dimensional control tasks — Cart Pole, Mountain Car and Lunar Lander, to difficult high-dimensional tasks — Inverted Double Pendulum, Hopper and HalfCheetah. Our results suggest that, the GAIRL agent can generally surpass the performance of demonstrations in both low-dimensional and high-dimensional tasks and get an optimal or close to optimal policy.

Deep reinforcement learning has shown promise in learning robust skills for robot control, but typically requires a large amount of samples to achieve good performance. Sim-to-real transfer learning has been developed to solve this problem, but the policy trained in simulation usually has unsatisfactory performance in the real world because simulators inevitably model the dynamics of reality imperfectly. To enable sample-efficient learning in the real world, we proposed progressive policy transfer with adaptive dynamics model (PPTADM). PPTADM assumes the dynamics of simulation and real world do not match but the state space is the same, transfers policy from simulation via progressive neural network (PNN) and further improves the policy with a learned forward dynamics model in reality. In addition, for real-world tasks in which reward functions are difficult or even impossible to define and verify the effectiveness, PPTADM can learn in real world solely from a transferred reward function that is estimated from simulation even though their dynamics do not match. Our results in five simulated tasks and on a real robot arm show that with PPTADM, the robot's learning efficiency and performance in the real world can be significantly improved.

Reinforcement learning (RL) can potentially be applied to real-world robot control in complex and uncertain environments. However, it is difficult or even unpractical to design an efficient reward function for various tasks, especially those large and high-dimensional environments. Generative adversarial imitation learning (GAIL) - a general model-free imitation learning method, allows robots to directly learn policies from expert trajectories in large and high-dimensional environments. However, GAIL is still sample inefficient in terms of environmental interaction. In this paper, to solve this problem, we propose a model-based adversarial imitation learning from demonstrations and human reward (MAILDH), a novel model-based interactive imitation framework combining the advantages of GAIL, interactive RL and model-based RL. We tested our method in eight physics-based discrete and continuous control tasks for RL. Our results show that MAILDH can greatly improve the sample efficiency and robustness compared to the original GAIL.

We propose a human-in-the-loop reinforcement learning mechanism to help robots learn emotional behavior. Unlike the previous methods of providing explicit feedback via pressing keyboard buttons or mouse clicks, we provide a more natural way for ordinary people to train social robots how to perform social tasks according to their preferences - facial expressions. The whole experiment is carried out on the desktop robot Haru, which is mainly used for the research of emotion and empathy participation. Our experimental results show that through learning from implicit feedback of facial features, Haru can quickly understand and dynamically adapt to individual preferences, and obtain a similar performance to learning from explicit feedback. In addition, we observe that the recognition error of human feedback will cause a “temporary regress” of the robot's learning performance, which is more obvious at the beginning of the training process. This phenomenon is shown to be correlated with the accuracy of recognizing negative implicit feedback.

Social robots that are able to express emotions can potentially improve human’s well-being. Whether and how they can learn from interactions between them and human being in a natural way will be key to their success and acceptance by ordinary people. In this paper, we proposed to shape social robot Haru affective behaviors with predicted continuous rewards based on received implicit facial feedback via human-centered reinforcement learning. The implicit facial feedback was estimated with the valence and arousal of received implicit facial feedback using Russell’s circumplex model, which can provide a more accurate estimation of the subtle psychological changes of human user, resulting in more effective robot behavior learning. The whole experiment is conducted on the desktop robot Haru, which is primarily used to study emotional interactions with human in different scenarios. Our experimental results show that with our proposed method, Haru can obtain a similar performance to learning from explicit feedback, eliminating the need for human users to get familiar with training interface in advance and resulting in an unobtrusive learning process.

The tabletop robot Haru, used for affective telepresence research, enables a teleoperator to communicate affects from a distance. The robot’s expressiveness offers myriad ways of communicating affects through the execution of emotive routines. The teleoperator reacts to input modalities such as the user’s facial expression, gestures and speech-based intent as perceived by the robot’s perception system. However, due to the sheer number of routines to select from, the task of choosing the appropriate or the most preferred routine is becoming cumbersome. In this paper, we propose a human-in-the-loop reinforcement learning mechanism in which an agent learns the teleoperator’s selection preference as a function of the input modalities and aids the routine selection process by narrowing it to n-best optimal choices. Our experimental results show that with only a few number of interactions from the teleoperator, the system can learn to recommend optimal routine behaviors for all perceived modalities, which greatly reduces the workload of the teleoperator.

Deep reinforcement learning has achieved significant success in many fields, but will confront sampling efficiency and safety problems when applying to robot control in the real world. Sim-to-real transfer learning was proposed to make use of samples in the simulation and overcome the gap between simulation and real world. In this paper, we focus on improving Progressive Neural Network — an effective sim-to-real learning method, by proposing Interactive Progressive Network Learning (IPNL). IPNL integrates progressive network and interactive reinforcement learning (interactive RL) which learns from evaluative feedback provided by an observing human trainer. We test our method using five RL tasks with discrete or continuous actions in OpenAI Gym and a sinusoids curve following task with AUV simulator on the Gazebo platform. Our results suggest that while Progressive Network has good performance when transferring from tasks with low-dimensional state space to those with high-dimensional one but has little effect for transferring from high-dimensional tasks to low-dimensional ones, IPNL allows an agent to learn a more stable policy with better performance faster for both cases. More importantly, our further analysis indicate that there is a synergy between Progressive Network and interactive RL for improving the agent’s learning. Our results in the path following of AUV shed light on the potential of applying our method in the real world tasks.

We describe a method of teaching a robot its empathic behavioural response from its interaction with people. We used the input modalities such as relative spatial information, facial expressions, body gestures and speech information as perception input that triggers the robot’s empathic response. First, we bootstrap the training through a pre-learning mechanism in which training is conducted by users who know the robotic system. This phase provides simulation-based training using a simple graphical user interface to simulate the input, rewards and correction feedback. In the second phase, we developed an online learning scheme for naive users to personalize their robot further, building on top of the bootstrapped model. Here, we developed a natural user interface that enables natural human-robot interaction via the suite of sensors that allows the users to provide evaluative feedback during the interaction with the robot. We evaluated the system and our results show that bootstrapping is an efficient tool to hasten the robot’s learning while online learning provided some form of personalization in the real environment with naive users.

As a branch of reinforcement learning, interactive reinforcement learning mainly studies the interaction process between humans and agents, allowing agents to learn from the intentions of human users and adapt to their preferences. In most of the current studies, human users need to intentionally provide explicit feedback via pressing keyboard buttons or mouse clicks. However, in our paper, we proposed an interactive reinforcement learning method that facilitates an agent to learn from human social signals - facial feedback via a ordinary camera and gestural feedback via a leap motion sensor. Our method provides a natural way for ordinary people to train agents how to perform a task according to their preferences. We tested our method in two reinforcement learning benchmarking domains - LoopMaze and Tetris, and compared to the state of the art - the TAMER framework. Our experimental results show that when learning from facial feedback the recognition of which is very low, the TAMER agent can get a similar performance to that of learning from keypress feedback with slightly more feedback. When learning from gestural feedback with a more accurate recognition, the TAMER agent can obtain a similar performance to that of learning from keypress feedback with much less feedback received. Moreover, our results indicate that the recognition error of facial feedback has a large effect on the agent performance in the beginning training process than in the later training stage. Finally, our results indicate that with enough recognition accuracy, human social signals can effectively improve the learning efficiency of agents with less human feedback.

Abstract Interactive reinforcement learning provides a way for agents to learn to solve tasks from evaluative feedback provided by a human user. Previous research showed that humans give copious feedback early in training but very sparsely thereafter. In this article, we investigate the potential of agent learning from trainers’ facial expressions via interpreting them as evaluative feedback. To do so, we implemented TAMER which is a popular interactive reinforcement learning method in a reinforcement-learning benchmark problem—Infinite Mario, and conducted the first large-scale study of TAMER involving 561 participants. With designed CNN–RNN model, our analysis shows that telling trainers to use facial expressions and competition can improve the accuracies for estimating positive and negative feedback using facial expressions. In addition, our results with a simulation experiment show that learning solely from predicted feedback based on facial expressions is possible and using strong/effective prediction models or a regression method, facial responses would significantly improve the performance of agents. Furthermore, our experiment supports previous studies demonstrating the importance of bi-directional feedback and competitive elements in the training interface.

Human-centered reinforcement learning (RL), in which an agent learns how to perform a task from evaluative feedback delivered by a human observer, has become more and more popular in recent years. The advantage of being able to learn from human feedback for a RL agent has led to increasing applicability to real-life problems. This paper describes the state-of-the-art human centered RL algorithms and aims to become a starting point for researchers who are initiating their endeavors in human-centered RL. Moreover, the objective of this paper is to present a comprehensive survey of the recent breakthroughs in this field and provide references to the most interesting and successful works. After starting with an introduction of the concepts of RL from environmental reward, this paper discusses the origins of human-centered RL and its difference from traditional RL. Then we describe different interpretations of human evaluative feedback, which have produced many human-centered RL algorithms in the past decade. In addition, we describe research on agents learning from both human evaluative feedback and environmental rewards as well as on improving the efficiency of human-centered RL. Finally, we conclude with an overview of application areas and a discussion of future work and open questions.

Programing robots to perform tasks is difficult in the real world because of its richness and uncertainty. For robots and agents to be more useful, they must be able to learn quickly from ordinary people via natural interactions. In this paper, we investigate how an agent can learn from demonstration and positive and negative evaluative feedback provided by a human teacher. Specifically, we proposed a model-based method-IRL-TAMER-by combining learning from demonstration via inverse reinforcement learning (IRL) and learning from human reward via the TAMER framework. We tested our method in the Grid World domain and compared with the TAMER framework using different discount factors on human reward. Our results suggest that although an agent learning via IRL can learn a useful value function indicating which states are good based on the demonstration, it cannot obtain an effective policy navigating to the goal state with one demonstration. However, learning from demonstration can reduce the number of human reward needed to obtain an optimal policy, especially the number of negative feedback. That is to say, learning from demonstration can be a jump-start for agent's learning from human reward and reduce the number of mistakes-incorrect actions. Furthermore, our results show that learning from demonstration can only be useful for agent's learning from human reward when the discount factor is small, i.e., learning from myopic human reward.

Learning from rewards generated by a human trainer observing an agent in action has been proven to be a powerful method for teaching autonomous agents to perform challenging tasks, especially for those non-technical users. Since the efficacy of this approach depends critically on the reward the trainer provides, we consider how the interaction between the trainer and the agent should be designed so as to increase the efficiency of the training process. This article investigates the influence of the agent’s socio-competitive feedback on the human trainer’s training behavior and the agent’s learning. The results of our user study with 85 participants suggest that the agent’s passive socio-competitive feedback—showing performance and score of agents trained by trainers in a leaderboard—substantially increases the engagement of the participants in the game task and improves the agents’ performance, even though the participants do not directly play the game but instead train the agent to do so. Moreover, making this feedback active—sending the trainer her agent’s performance relative to others—further induces more participants to train agents longer and improves the agent’s learning. Our further analysis shows that agents trained by trainers affected by both the passive and active social feedback could obtain a higher performance under a score mechanism that could be optimized from the trainer’s perspective and the agent’s additional active social feedback can keep participants to further train agents to learn policies that can obtain a higher performance under such a score mechanism.

In this work, we address a relatively unexplored aspect of designing agents that learn from human reward. We investigate how an agent’s non-task behavior can affect a human trainer’s training and agent learning. We use the TAMER framework, which facilitates the training of agents by human-generated reward signals, i.e., judgements of the quality of the agent’s actions, as the foundation for our investigation. Then, starting from the premise that the interaction between the agent and the trainer should be bi-directional, we propose two new training interfaces to increase a human trainer’s active involvement in the training process and thereby improve the agent’s task performance. One provides information on the agent’s uncertainty which is a metric calculated as data coverage, the other on its performance. Our results from a 51-subject user study show that these interfaces can induce the trainers to train longer and give more feedback. The agent’s performance, however, increases only in response to the addition of performance-oriented information, not by sharing uncertainty levels. These results suggest that the organizational maxim about human behavior, “you get what you measure”—i.e., sharing metrics with people causes them to focus on optimizing those metrics while de-emphasizing other objectives—also applies to the training of agents. Using principle component analysis, we show how trainers in the two conditions train agents differently. In addition, by simulating the influence of the agent’s uncertainty–informative behavior on a human’s training behavior, we show that trainers could be distracted by the agent sharing its uncertainty levels about its actions, giving poor feedback for the sake of reducing the agent’s uncertainty without improving the agent’s performance.