Reinforcement learning (RL) has emerged as a promising paradigm in complex and continuous robotic tasks, however, safe exploration has been one of the main challenges, especially in contact-rich manipulation tasks in unstructured environments. Focusing on this issue, we propose SRL-VIC: a model-free safe RL framework combined with a variable impedance controller (VIC). Specifically, safety critic and recovery policy networks are pre-trained where safety critic evaluates the safety of the next action using a risk value before it is executed and the recovery policy suggests a corrective action if the risk value is high. Furthermore, the policies are updated online where the task policy not only achieves the task but also modulates the stiffness parameters to keep a safe and compliant profile. A set of experiments in contact-rich maze tasks demonstrate that our framework outperforms the baselines (without the recovery mechanism and without the VIC), yielding a good trade-off between efficient task accomplishment and safety guarantee. We show our policy trained on simulation can be deployed on a physical robot without fine-tuning, achieving successful task completion with robustness and generalization. The video is available at https://youtu.be/ksWXR3vByoQ.

Aligning human preference and value is an important requirement for building contemporary foundation models and embodied AI. However, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into successive stages, such as supervised fine-tuning (SFT), reward modeling (RM), and reinforcement learning (RL), each performing one specific learning task. Such a sequential approach results in serious issues such as significant under-utilization of data and distribution mismatch between the learned reward model and generated policy, which eventually lead to poor alignment performance. We develop a single stage approach named Alignment with Integrated Human Feedback (AIHF), capable of integrating both human preference and demonstration to train reward models and the policy. The proposed approach admits a suite of efficient algorithms, which can easily reduce to, and leverage, popular alignment algorithms such as RLHF and Directly Policy Optimization (DPO), and only requires minor changes to the existing alignment pipelines. We demonstrate the efficiency of the proposed solutions with extensive experiments involving alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithms such as RLHF and DPO by large margins, especially when the amount of high-quality preference data is relatively limited.

Abstract-- This paper employs a reinforcement learningbased model identification method aimed at enhancing the accuracy of the dynamics for our snake robot, called COBRA. Leveraging gradient information and iterative optimization, the proposed approach refines the parameters of COBRA's dynamical model such as coefficient of friction and actuator parameters using experimental and simulated data. Experimental validation on the hardware platform demonstrates the efficacy of the proposed approach, highlighting its potential to address simto-real gap in robot implementation. These systems present formidable challenges in modeling and control due to the complex interplay of unilateral contact forces, leading to intricate complementarity conditions [2]. Traditional approaches have previously tackled these force inclusion issues with promising outcomes [3], [4].

Effective multi-intersection collaboration is pivotal for reinforcement-learning-based traffic signal control to alleviate congestion. Existing work mainly chooses neighboring intersections as collaborators. However, quite an amount of congestion, even some wide-range congestion, is caused by non-neighbors failing to collaborate. To address these issues, we propose to separate the collaborator selection as a second policy to be learned, concurrently being updated with the original signal-controlling policy. Specifically, the selection policy in real-time adaptively selects the best teammates according to phase- and intersection-level features. Empirical results on both synthetic and real-world datasets provide robust validation for the superiority of our approach, offering significant improvements over existing state-of-the-art methods. The code is available at https://github.com/bonaldli/CoSLight.

Sample efficiency is critical when applying learning-based methods to robotic manipulation due to the high cost of collecting expert demonstrations and the challenges of on-robot policy learning through online Reinforcement Learning (RL). Offline RL addresses this issue by enabling policy learning from an offline dataset collected using any behavioral policy, regardless of its quality. However, recent advancements in offline RL have predominantly focused on learning from large datasets. Given that many robotic manipulation tasks can be formulated as rotation-symmetric problems, we investigate the use of $SO(2)$-equivariant neural networks for offline RL with a limited number of demonstrations. Our experimental results show that equivariant versions of Conservative Q-Learning (CQL) and Implicit Q-Learning (IQL) outperform their non-equivariant counterparts. We provide empirical evidence demonstrating how equivariance improves offline learning algorithms in the low-data regime.

Sample efficiency is a critical challenge in Reinforcement Learning. Model-based RL has emerged as a solution, but its application has largely been confined to single-agent scenarios. In this work, we introduce CoDreamer, an extension of the Dreamer algorithm for multi-agent environments. CoDreamer leverages Graph Neural Networks for a two-level communication system to tackle challenges such as partial observability and inter-agent cooperation. Communication is separately utilised within the learned world models and within the learned policies of each agent to enhance modelling and task-solving. We show that CoDreamer offers greater expressive power than a naive application of Dreamer, and we demonstrate its superiority over baseline methods across various multi-agent environments.

It is crucial that users are empowered to take advantage of the functionality of a robot and use their understanding of that functionality to perform novel and creative tasks. Given a robot trained with Reinforcement Learning (RL), a user may wish to leverage that autonomy along with their familiarity of how they expect the robot to behave to collaborate with the robot. One technique is for the user to take control of some of the robot's action space through teleoperation, allowing the RL policy to simultaneously control the rest. We formalize this type of shared control as Partitioned Control (PC). However, this may not be possible using an out-of-the-box RL policy. For example, a user's control may bring the robot into a failure state from the policy's perspective, causing it to act unexpectedly and hindering the success of the user's desired task. In this work, we formalize this problem and present Imaginary Out-of-Distribution Actions, IODA, an initial algorithm which empowers users to leverage their expectations of a robot's behavior to accomplish new tasks. We deploy IODA in a user study with a real robot and find that IODA leads to both better task performance and a higher degree of alignment between robot behavior and user expectation. We also show that in PC, there is a strong and significant correlation between task performance and the robot's ability to meet user expectations, highlighting the need for approaches like IODA. Code is available at https://github.com/AABL-Lab/ioda_roman_2024

Recent advancements in off-policy Reinforcement Learning (RL) have significantly improved sample efficiency, primarily due to the incorporation of various forms of regularization that enable more gradient update steps than traditional agents. However, many of these techniques have been tested in limited settings, often on tasks from single simulation benchmarks and against well-known algorithms rather than a range of regularization approaches. This limits our understanding of the specific mechanisms driving RL improvements. To address this, we implemented over 60 different off-policy agents, each integrating established regularization techniques from recent state-of-the-art algorithms. We tested these agents across 14 diverse tasks from 2 simulation benchmarks, measuring training metrics related to overestimation, overfitting, and plasticity loss -- issues that motivate the examined regularization techniques. Our findings reveal that while the effectiveness of a specific regularization setup varies with the task, certain combinations consistently demonstrate robust and superior performance. Notably, a simple Soft Actor-Critic agent, appropriately regularized, reliably finds a better-performing policy within the training regime, which previously was achieved mainly through model-based approaches.

Coping with intensively interactive scenarios is one of the significant challenges in the development of autonomous driving. Reinforcement learning (RL) offers an ideal solution for such scenarios through its self-evolution mechanism via interaction with the environment. However, the lack of sufficient safety mechanisms in common RL leads to the fact that agent often find it difficult to interact well in highly dynamic environment and may collide in pursuit of short-term rewards. Much of the existing safe RL methods require environment modeling to generate reliable safety boundaries that constrain agent behavior. Nevertheless, acquiring such safety boundaries is not always feasible in dynamic environments. Inspired by the driver's behavior of acting when uncertainty is minimal, this study introduces the concept of action timing to replace explicit safety boundary modeling. We define "actor" as an agent to decide optimal action at each step. By imaging the actor take opportunity to act as a timing-dependent gradual process, the other agent called "timing taker" can evaluate the optimal action execution time, and relate the optimal timing to each action moment as a dynamic safety factor to constrain the actor's action. In the experiment involving a complex, unsignaled intersection interaction, this framework achieved superior safety performance compared to all benchmark models.

One of the most natural approaches to reinforcement learning (RL) with function approximation is value iteration, which inductively generates approximations to the optimal value function by solving a sequence of regression problems. To ensure the success of value iteration, it is typically assumed that Bellman completeness holds, which ensures that these regression problems are well-specified. We study the problem of learning an optimal policy under Bellman completeness in the online model of RL with linear function approximation. In the linear setting, while statistically efficient algorithms are known under Bellman completeness (e.g., Jiang et al. (2017); Zanette et al. (2020)), these algorithms all rely on the principle of global optimism which requires solving a nonconvex optimization problem. In particular, it has remained open as to whether computationally efficient algorithms exist. In this paper we give the first polynomial-time algorithm for RL under linear Bellman completeness when the number of actions is any constant.