Humans excel at reusing prior knowledge to address new challenges and developing skills while solving problems. This paradigm becomes increasingly popular in the development of autonomous agents, as it develops systems that can self-evolve in response to new challenges like human beings. However, previous methods suffer from limited training efficiency when expanding new skills and fail to fully leverage prior knowledge to facilitate new task learning. We propose Parametric Skill Expansion and Composition (PSEC), a new framework designed to iteratively evolve the agents' capabilities and efficiently address new challenges by maintaining a manageable skill library. This library can progressively integrate skill primitives as plug-and-play Low-Rank Adaptation (LoRA) modules in parameter-efficient finetuning, facilitating efficient and flexible skill expansion. This structure also enables the direct skill compositions in parameter space by merging LoRA modules that encode different skills, leveraging shared information across skills to effectively program new skills. Based on this, we propose a context-aware modular to dynamically activate different skills to collaboratively handle new tasks. Empowering diverse applications including multi-objective composition, dynamics shift, and continual policy shift, the results on D4RL, DSRL benchmarks, and the DeepMind Control Suite show that PSEC exhibits superior capacity to leverage prior knowledge to efficiently tackle new challenges, as well as expand its skill libraries to evolve the capabilities. Project website: https://ltlhuuu.github.io/PSEC/.

In offline reinforcement learning, value overestimation caused by out-of-distribution (OOD) actions significantly limits policy performance. Recently, diffusion models have been leveraged for their strong distribution-matching capabilities, enforcing conservatism through behavior policy constraints. However, existing methods often apply indiscriminate regularization to redundant actions in low-quality datasets, resulting in excessive conservatism and an imbalance between the expressiveness and efficiency of diffusion modeling. To address these issues, we propose DIffusion policies with Value-conditional Optimization (DIVO), a novel approach that leverages diffusion models to generate high-quality, broadly covered in-distribution state-action samples while facilitating efficient policy improvement. Specifically, DIVO introduces a binary-weighted mechanism that utilizes the advantage values of actions in the offline dataset to guide diffusion model training. This enables a more precise alignment with the dataset’s distribution while selectively expanding the boundaries of high-advantage actions. During policy improvement, DIVO dynamically filters high-return-potential actions from the diffusion model, effectively guiding the learned policy toward better performance. This approach achieves a critical balance between conservatism and explorability in offline RL. We evaluate DIVO on the D4RL benchmark and compare it against state-of-the-art baselines. Empirical results demonstrate that DIVO achieves superior performance, delivering significant improvements in average returns across locomotion tasks and outperforming existing methods in the challenging AntMaze domain, where sparse rewards pose a major difficulty.

The agile maneuvering control of autonomous vehicles (AVs) requires the tracking of reference trajectories characterized by high acceleration, sharp curvature, considerable disturbances, and significant time-varying, all while ensuring stability and accuracy. The inherent uncertainty and time-varying nature of both the vehicle model and its environment pose significant challenges to achieving high-performance tracking during agile maneuvers. Developing a control algorithm that enables solving the optimal policy for nonlinear systems with uncertainties is critical. In this paper, we propose a learning-based predictive control approach, namely, an adaptive model predictive control (AMPC) with Actor-Critic Learning (ACL) for generating closed-loop MPC policies for agile maneuvering of AVs. The proposed approach leverages neural networks to model the dynamics uncertainties online. The control policy and model are updated simultaneously to realize performance op-timization under time-varying uncertainties. Simulation results demonstrate that our proposed algorithm outperforms other leading ACL methods, as well as MPC and Linear Quadratic Regulator (LQR). Furthermore, field test experiment results validate its effectiveness on the HongQi-EHS3 electric vehicle, showing superior control performance compared to MPC both on paved roads and curved off-roads with excellent stability performance.

Safe reinforcement learning (RL) aims to optimize the task performance with safety guarantees. One common modeling scheme to study safe RL problems is the constrained Markov decision process (CMDP). However, current safe RL methods within the CMDP framework face challenges in tradeoffs among various objectives and gradient conflicts of policy updating. To cope with these challenges, this article presents a novel safe RL approach called game-theoretic constrained policy optimization (GCPO). The proposed approach formulates the CMDP problem as a general-sum Markov game with multiple players, where a task player seeks to maximize the reward objective, while constraint players aim to minimize constraint objectives until they are fulfilled. By doing so, GCPO adopts the learning mode with multiple subpolicies, each aligned with a distinct objective, that collectively constitute the overall behavior of the agent. The learning convergence of the GCPO can be ensured with the contraction mapping to the Nash equilibrium. Furthermore, a novel dominant timescale update rule is presented for multiplayer policy learning to guarantee constraint satisfaction. The learning convergence and constraint satisfaction of GCPO are theoretically analyzed. Consequently, GCPO eliminates the necessity of tuning tradeoff parameters and mitigates gradient conflicts during multiobjective policy updating. Experimental results show that GCPO outperforms state-of-the-art safe RL algorithms in a quadrotor trajectory tracking task and various high-dimensional robot locomotion benchmarks. Moreover, GCPO exhibits robustness to diverse scales of task rewards and constraint costs without the need for intricate tradeoffs.

In offline reinforcement learning, the challenge of out-of-distribution (OOD) is pronounced. To address this, existing methods often constrain the learned policy through policy regularization. However, these methods often suffer from the issue of unnecessary conservativeness, hampering policy improvement. This occurs due to the indiscriminate use of all actions from the behavior policy that generates the offline dataset as constraints. The problem becomes particularly noticeable when the quality of the dataset is suboptimal. Thus, we propose Adaptive Advantage-guided Policy Regularization (A2PR), obtaining high-advantage actions from an augmented behavior policy combined with VAE to guide the learned policy. A2PR can select high-advantage actions that differ from those present in the dataset, while still effectively maintaining conservatism from OOD actions. This is achieved by harnessing the VAE capacity to generate samples matching the distribution of the data points. We theoretically prove that the improvement of the behavior policy is guaranteed. Besides, it effectively mitigates value overestimation with a bounded performance gap. Empirically, we conduct a series of experiments on the D4RL benchmark, where A2PR demonstrates state-of-the-art performance. Furthermore, experimental results on additional suboptimal mixed datasets reveal that A2PR exhibits superior performance. Code is available at https://github.com/ltlhuuu/A2PR.

In recent years, the multiple traveling salesmen problem (MTSP or multiple TSP) has received increasing research interest and one of its main applications is coordinated multirobot mission planning, such as cooperative search and rescue tasks. However, it is still challenging to solve MTSP with improved inference efficiency as well as solution quality in varying situations, e.g., different city positions, different numbers of cities, or agents. In this article, we propose an attention-based multiagent reinforcement learning (AMARL) approach, which is based on the gated transformer feature representations for min-max multiple TSPs. The state feature extraction network in our proposed approach adopts the gated transformer architecture with reordering layer normalization (LN) and a new gate mechanism. It aggregates fixed-dimensional attention-based state features irrespective of the number of agents and cities. The action space of our proposed approach is designed to decouple the interaction of agents' simultaneous decision-making. At each time step, only one agent is assigned to a non-zero action so that the action selection strategy can be transferred across tasks with different numbers of agents and cities. Extensive experiments on min-max multiple TSPs were conducted to illustrate the effectiveness and advantages of the proposed approach. Compared with six representative algorithms, our proposed approach achieves state-of-the-art performance in solution quality and inference efficiency. In particular, the proposed approach is suitable for tasks with different numbers of agents or cities without extra learning, and experimental results demonstrate that the proposed approach realizes powerful transfer capability across tasks.

Deep reinforcement learning (RL) typically requires a tremendous number of training samples, which are not practical in many applications. State abstraction and world models are two promising approaches for improving sample efficiency in deep RL. However, both state abstraction and world models may degrade the learning performance. In this article, we propose an abstracted model-based policy learning (AMPL) algorithm, which improves the sample efficiency of deep RL. In AMPL, a novel state abstraction method via multistep bisimulation is first developed to learn task-related latent state spaces. Hence, the original Markov decision processes (MDPs) are compressed into abstracted MDPs. Then, a causal transformer model predictor (CTMP) is designed to approximate the abstracted MDPs and generate long-horizon simulated trajectories with a smaller multistep prediction error. Policies are efficiently learned through these trajectories within the abstracted MDPs via a modified multistep soft actor-critic algorithm with a λ -target. Moreover, theoretical analysis shows that the AMPL algorithm can improve sample efficiency during the training process. On Atari games and the DeepMind Control (DMControl) suite, AMPL surpasses current state-of-the-art deep RL algorithms in terms of sample efficiency. Furthermore, DMControl tasks with moving noises are conducted, and the results demonstrate that AMPL is robust to task-irrelevant observational distractors and significantly outperforms the existing approaches.