Multimodal Language Models have gained significant traction for their ability to process diverse input data types and generate coherent, contextually relevant outputs across various applications. While supervised fine-tuning (SFT) has been the predominant approach to enhance MLLM capabilities in task-specific optimization, it often falls short in fostering crucial generalized reasoning abilities. Despite the potential of reinforcement learning (RL) to address these limitations, it faces two issues: (1) its generalized capabilities in multimodal tasks remain underexplored. (2) its training constraints such as constant Kullback-Leibler or clamp strategy easily lead to suboptimal bottleneck. To adress these issues, we introduce OThink-MR1, a framework that extends RL to MLLMs, enabling them to achieve deeper understanding and reasoning across multimodal tasks. We design a dynamic Kullback-Leibler strategy that significantly enhances RL performance, surpassing SFT in same-task evaluations. Also, we are the first to reveal that RL exhibits remarkable cross-task generalization capabilities, which shows that models post-trained with RL on one multimodal task can be effectively transfered to another tasks. Finally, extensive experiments demonstrate the great reasoning ability of our proposed OThink-MR1.

Achieving controlled jumping behaviour for a quadruped robot is a challenging task, especially when introducing passive compliance in mechanical design. This study addresses this challenge via imitation-based deep reinforcement learning with a progressive training process. To start, we learn the jumping skill by mimicking a coarse jumping example generated by model-based trajectory optimization. Subsequently, we generalize the learned policy to broader situations, including various distances in both forward and lateral directions, and then pursue robust jumping in unknown ground unevenness. In addition, without tuning the reward much, we learn the jumping policy for a quadruped with parallel elasticity. Results show that using the proposed method, i) the robot learns versatile jumps by learning only from a single demonstration, ii) the robot with parallel compliance reduces the landing error by 11.1%, saves energy cost by 15.2% and reduces the peak torque by 15.8%, compared to the rigid robot without parallel elasticity, iii) the robot can perform jumps of variable distances with robustness against ground unevenness (maximal 4cm height perturbations) using only proprioceptive perception.

This work concerns itself with the task of reconstructing all edges of an arbitrary 3D wire-frame model projected to an image plane. We explore a bottom-up part-wise procedure undertaken by an RL agent to segment and reconstruct these 2D multipart objects. The environment's state is represented as a four-colour image, where different colours correspond to background, a target edge, a reconstruction line, and the overlap of both. At each step, the agent can transform the reconstruction line within a four-dimensional action space or terminate the episode using a specific termination action. To investigate the impact of reward function formulations, we tested episodic and incremental rewards, as well as combined approaches. Empirical results demonstrated that the latter yielded the most effective training performance. To further enhance efficiency and stability, we introduce curriculum learning strategies. First, an action-based curriculum was implemented, where the agent was initially restricted to a reduced action space, being able to only perform three of the five possible actions, before progressing to the full action space. Second, we test a task-based curriculum, where the agent first solves a simplified version of the problem before being presented with the full, more complex task. This second approach produced promising results, as the agent not only successfully transitioned from learning the simplified task to mastering the full task, but in doing so gained significant performance. This study demonstrates the potential of an iterative RL wire-frame reconstruction in two dimensions. By combining optimized reward function formulations with curriculum learning strategies, we achieved significant improvements in training success. The proposed methodology provides an effective framework for solving similar tasks and represents a promising direction for future research in the field.

In this paper, we propose Unreal Multi-Agent Playground (Unreal-MAP), an MARL general platform based on the Unreal-Engine (UE). Unreal-MAP allows users to freely create multi-agent tasks using the vast visual and physical resources available in the UE community, and deploy state-of-the-art (SOTA) MARL algorithms within them. Unreal-MAP is user-friendly in terms of deployment, modification, and visualization, and all its components are open-source. We also develop an experimental framework compatible with algorithms ranging from rule-based to learning-based provided by third-party frameworks. Lastly, we deploy several SOTA algorithms in example tasks developed via Unreal-MAP, and conduct corresponding experimental analyses. We believe Unreal-MAP can play an important role in the MARL field by closely integrating existing algorithms with user-customized tasks, thus advancing the field of MARL.

In this work, we study the sample complexity problem of risk-sensitive Reinforcement Learning (RL) with a generative model, where we aim to maximize the Conditional Value at Risk (CVaR) with risk tolerance level $\tau$ at each step, named Iterated CVaR. We first build a connection between Iterated CVaR RL with $(s, a)$-rectangular distributional robust RL with the specific uncertainty set for CVaR. We develop nearly matching upper and lower bounds on the sample complexity for this problem. Specifically, we first prove that a value iteration-based algorithm, ICVaR-VI, achieves an $\epsilon$-optimal policy with at most $\overset{\sim}{O}\left(\frac{SA}{(1-\gamma)^4\tau^2\epsilon^2}\right)$ samples, where $\gamma$ is the discount factor, and $S, A$ are the sizes of the state and action spaces. Furthermore, if $\tau \geq \gamma$, then the sample complexity can be further improved to $\overset{\sim}{O}\left( \frac{SA}{(1-\gamma)^3\epsilon^2} \right)$. We further show a minimax lower bound of $\overset{\sim}{O} \left(\frac{(1-\gamma \tau)SA}{(1-\gamma)^4\tau\epsilon^2}\right)$. For a constant risk level $0<\tau\leq 1$, our upper and lower bounds match with each other, demonstrating the tightness and optimality of our analyses.We also investigate a limiting case with a small risk level $\tau$, called Worst-Path RL, where the objective is to maximize the minimum possible cumulative reward. We develop matching upper and lower bounds of $\overset{\sim}{O}\left(\frac{SA}{p_{\min}}\right)$, where $p_{\min}$ denotes the minimum non-zero reaching probability of the transition kernel.

A pervasive challenge in Reinforcement Learning (RL) is the "curse of dimensionality" which is the exponential growth in the state-action space when optimizing a high-dimensional target task. The framework of curriculum learning trains the agent in a curriculum composed of a sequence of related and more manageable source tasks. The expectation is that when some optimal decision rules are shared across source tasks and the target task, the agent could more quickly pick up the necessary skills to behave optimally in the environment, thus accelerating the learning process. However, this critical assumption of invariant optimal decision rules does not necessarily hold in many practical applications, specifically when the underlying environment contains unobserved confounders. This paper studies the problem of curriculum RL through causal lenses. We derive a sufficient graphical condition characterizing causally aligned source tasks, i.e., the invariance of optimal decision rules holds. We further develop an efficient algorithm to generate a causally aligned curriculum, provided with qualitative causal knowledge of the target task. Finally, we validate our proposed methodology through experiments in discrete and continuous confounded tasks with pixel observations.

Soft robotics holds transformative potential for enabling adaptive and adaptable systems in dynamic environments. However, the interplay between morphological and control complexities and their collective impact on task performance remains poorly understood. Therefore, in this study, we investigate these trade-offs across tasks of differing difficulty levels using four well-used morphological complexity metrics and control complexity measured by FLOPs. We investigate how these factors jointly influence task performance by utilizing the evolutionary robot experiments. Results show that optimal performance depends on the alignment between morphology and control: simpler morphologies and lightweight controllers suffice for easier tasks, while harder tasks demand higher complexities in both dimensions. In addition, a clear trade-off between morphological and control complexities that achieve the same task performance can be observed. Moreover, we also propose a sensitivity analysis to expose the task-specific contributions of individual morphological metrics. Our study establishes a framework for investigating the relationships between morphology, control, and task performance, advancing the development of task-specific robotic designs that balance computational efficiency with adaptability. This study contributes to the practical application of soft robotics in real-world scenarios by providing actionable insights.

This paper presents CONTHER, a novel reinforcement learning algorithm designed to efficiently and rapidly train robotic agents for goal-oriented manipulation tasks and obstacle avoidance. The algorithm uses a modified replay buffer inspired by the Hindsight Experience Replay (HER) approach to artificially populate experience with successful trajectories, effectively addressing the problem of sparse reward scenarios and eliminating the need to manually collect expert demonstrations. The developed algorithm proposes a Transformer-based architecture to incorporate the context of previous states, allowing the agent to perform a deeper analysis and make decisions in a manner more akin to human learning. The effectiveness of the built-in replay buffer, which acts as an "internal demonstrator", is twofold: it accelerates learning and allows the algorithm to adapt to different tasks. Empirical data confirm the superiority of the algorithm by an average of 38.46% over other considered methods, and the most successful baseline by 28.21%, showing higher success rates and faster convergence in the point-reaching task. Since the control is performed through the robot's joints, the algorithm facilitates potential adaptation to a real robot system and construction of an obstacle avoidance task. Therefore, the algorithm has also been tested on tasks requiring following a complex dynamic trajectory and obstacle avoidance. The design of the algorithm ensures its applicability to a wide range of goal-oriented tasks, making it an easily integrated solution for real-world robotics applications.

This paper introduces novel Bellman mappings (B-Maps) for value iteration (VI) in distributed reinforcement learning (DRL), where multiple agents operate over a network without a centralized fusion node. Each agent constructs its own nonparametric B-Map for VI while communicating only with direct neighbors to achieve consensus. These B-Maps operate on Q-functions represented in a reproducing kernel Hilbert space, enabling a nonparametric formulation that allows for flexible, agent-specific basis function design. Unlike existing DRL methods that restrict information exchange to Q-function estimates, the proposed framework also enables agents to share basis information in the form of covariance matrices, capturing additional structural details. A theoretical analysis establishes linear convergence rates for both Q-function and covariance-matrix estimates toward their consensus values. The optimal learning rates for consensus-based updates are dictated by the ratio of the smallest positive eigenvalue to the largest one of the network's Laplacian matrix. Furthermore, each nodal Q-function estimate is shown to lie very close to the fixed point of a centralized nonparametric B-Map, effectively allowing the proposed DRL design to approximate the performance of a centralized fusion center. Numerical experiments on two well-known control problems demonstrate the superior performance of the proposed nonparametric B-Maps compared to prior methods. Notably, the results reveal a counter-intuitive finding: although the proposed approach involves greater information exchange -- specifically through the sharing of covariance matrices -- it achieves the desired performance with lower cumulative communication cost than existing DRL schemes, highlighting the crucial role of basis information in accelerating the learning process.

In many practical reinforcement learning scenarios, feedback is provided only at the end of a long horizon, leading to sparse and delayed rewards. Existing reward redistribution methods typically assume that per-step rewards are independent, thus overlooking interdependencies among state--action pairs. In this paper, we propose a \emph{Likelihood Reward Redistribution} (LRR) framework that addresses this issue by modeling each per-step reward with a parametric probability distribution whose parameters depend on the state--action pair. By maximizing the likelihood of the observed episodic return via a leave-one-out (LOO) strategy that leverages the entire trajectory, our framework inherently introduces an uncertainty regularization term into the surrogate objective. Moreover, we show that the conventional mean squared error (MSE) loss for reward redistribution emerges as a special case of our likelihood framework when the uncertainty is fixed under the Gaussian distribution. When integrated with an off-policy algorithm such as Soft Actor-Critic, LRR yields dense and informative reward signals, resulting in superior sample efficiency and policy performance on Box-2d and MuJoCo benchmarks.