Humans possess delicate dynamic balance mechanisms that enable them to maintain stability across diverse terrains and under extreme conditions. However, despite significant advances recently, existing locomotion algorithms for humanoid robots are still struggle to traverse extreme environments, especially in cases that lack external perception (e.g., vision or LiDAR). This is because current methods often rely on gait-based or perception-condition rewards, lacking effective mechanisms to handle unobservable obstacles and sudden balance loss. To address this challenge, we propose a novel whole-body locomotion algorithm based on dynamic balance and Reinforcement Learning (RL) that enables humanoid robots to traverse extreme terrains, particularly narrow pathways and unexpected obstacles, using only proprioception. Specifically, we introduce a dynamic balance mechanism by leveraging an extended measure of Zero-Moment Point (ZMP)-driven rewards and task-driven rewards in a whole-body actor-critic framework, aiming to achieve coordinated actions of the upper and lower limbs for robust locomotion. Experiments conducted on a full-sized Unitree H1-2 robot verify the ability of our method to maintain balance on extremely narrow terrains and under external disturbances, demonstrating its effectiveness in enhancing the robot's adaptability to complex environments. The videos are given at https://whole-body-loco.github.io.

In cooperative multi-agent reinforcement learning (MARL), well-designed communication protocols can effectively facilitate consensus among agents, thereby enhancing task performance. Moreover, in large-scale multi-agent systems commonly found in real-world applications, effective communication plays an even more critical role due to the escalated challenge of partial observability compared to smaller-scale setups. In this work, we endeavor to develop a scalable communication protocol for MARL. Unlike previous methods that focus on selecting optimal pairwise communication links-a task that becomes increasingly complex as the number of agents grows-we adopt a global perspective on communication topology design. Specifically, we propose utilizing the exponential topology to enable rapid information dissemination among agents by leveraging its small-diameter and small-size properties. This approach leads to a scalable communication protocol, named ExpoComm. To fully unlock the potential of exponential graphs as communication topologies, we employ memory-based message processors and auxiliary tasks to ground messages, ensuring that they reflect global information and benefit decision-making. Extensive experiments on large-scale cooperative benchmarks, including MAgent and Infrastructure Management Planning, demonstrate the superior performance and robust zero-shot transferability of ExpoComm compared to existing communication strategies. The code is publicly available at https://github.com/LXXXXR/ExpoComm.

Reward engineering is one of the key challenges in Reinforcement Learning (RL). Preference-based RL effectively addresses this issue by learning from human feedback. However, it is both time-consuming and expensive to collect human preference labels. In this paper, we propose a novel \textbf{V}ision-\textbf{L}anguage \textbf{P}reference learning framework, named \textbf{VLP}, which learns a vision-language preference model to provide preference feedback for embodied manipulation tasks. To achieve this, we define three types of language-conditioned preferences and construct a vision-language preference dataset, which contains versatile implicit preference orders without human annotations. The preference model learns to extract language-related features, and then serves as a preference annotator in various downstream tasks. The policy can be learned according to the annotated preferences via reward learning or direct policy optimization. Extensive empirical results on simulated embodied manipulation tasks demonstrate that our method provides accurate preferences and generalizes to unseen tasks and unseen language instructions, outperforming the baselines by a large margin.

Reinforcement Learning from Human Feedback (RLHF) has shown great potential in fine-tuning Large Language Models (LLMs) to align with human preferences. Existing methods perform preference alignment from a fixed dataset, which can be limited in data coverage, and the resulting reward model is hard to generalize in out-of-distribution responses. Thus, online RLHF is more desirable to empower the LLM to explore outside the support of the initial dataset by iteratively collecting the prompt-response pairs. In this paper, we study the fundamental problem in online RLHF, i.e. \emph{how to explore} for LLM. We give a theoretical motivation in linear reward assumption to show that an optimistic reward with an upper confidence bound (UCB) term leads to a provably efficient RLHF policy. Then, we reformulate our objective to direct preference optimization with an exploration term, where the UCB-term can be converted to a count-based exploration bonus. We further propose a practical algorithm, named \emph{Count-based Online Preference Optimization (COPO)}, which leverages a simple coin-flip counting module to estimate the pseudo-count of a prompt-response pair in previously collected data. COPO encourages LLMs to balance exploration and preference optimization in an iterative manner, which enlarges the exploration space and the entire data coverage of iterative LLM policies. We conduct online RLHF experiments on Zephyr and Llama-3 models. The results on instruction-following and standard academic benchmarks show that COPO significantly increases performance.

Automatic Radiology Report Generation (RRG) is an important topic for alleviating the substantial workload of radiologists. Existing RRG approaches rely on supervised regression based on different architectures or additional knowledge injection,while the generated report may not align optimally with radiologists' preferences. Especially, since the preferences of radiologists are inherently heterogeneous and multidimensional, e.g., some may prioritize report fluency, while others emphasize clinical accuracy. To address this problem,we propose a new RRG method via Multi-objective Preference Optimization (MPO) to align the pre-trained RRG model with multiple human preferences, which can be formulated by multi-dimensional reward functions and optimized by multi-objective reinforcement learning (RL). Specifically, we use a preference vector to represent the weight of preferences and use it as a condition for the RRG model. Then, a linearly weighed reward is obtained via a dot product between the preference vector and multi-dimensional reward. Next,the RRG model is optimized to align with the preference vector by optimizing such a reward via RL. In the training stage,we randomly sample diverse preference vectors from the preference space and align the model by optimizing the weighted multi-objective rewards, which leads to an optimal policy on the entire preference space. When inference,our model can generate reports aligned with specific preferences without further fine-tuning. Extensive experiments on two public datasets show the proposed method can generate reports that cater to different preferences in a single model and achieve state-of-the-art performance.

We consider off-dynamics reinforcement learning (RL) where one needs to transfer policies across different domains with dynamics mismatch. Despite the focus on developing dynamics-aware algorithms, this field is hindered due to the lack of a standard benchmark. To bridge this gap, we introduce ODRL, the first benchmark tailored for evaluating off-dynamics RL methods. ODRL contains four experimental settings where the source and target domains can be either online or offline, and provides diverse tasks and a broad spectrum of dynamics shifts, making it a reliable platform to comprehensively evaluate the agent's adaptation ability to the target domain. Furthermore, ODRL includes recent off-dynamics RL algorithms in a unified framework and introduces some extra baselines for different settings, all implemented in a single-file manner. To unpack the true adaptation capability of existing methods, we conduct extensive benchmarking experiments, which show that no method has universal advantages across varied dynamics shifts. We hope this benchmark can serve as a cornerstone for future research endeavors.

Diffusion models demonstrate superior performance in capturing complex distributions from large-scale datasets, providing a promising solution for quadrupedal locomotion control. However, offline policy is sensitive to Out-of-Distribution (OOD) states due to the limited state coverage in the datasets. In this work, we propose a two-stage learning framework combining offline learning and online preference alignment for legged locomotion control. Through the offline stage, the diffusion planner learns the joint distribution of state-action sequences from expert datasets without using reward labels. Subsequently, we perform the online interaction in the simulation environment based on the trained offline planer, which significantly addresses the OOD issues and improves the robustness. Specifically, we propose a novel weak preference labeling method without the ground-truth reward or human preferences. The proposed method exhibits superior stability and velocity tracking accuracy in pacing, trotting, and bounding gait under both slow- and high-speed scenarios and can perform zero-shot transfer to the real Unitree Go1 robots. The project website for this paper is at https://shangjaven.github.io/preference-aligned-diffusion-legged/.

Diffusion models have demonstrated their capabilities in modeling trajectories of multi-tasks. However, existing multi-task planners or policies typically rely on task-specific demonstrations via multi-task imitation, or require task-specific reward labels to facilitate policy optimization via Reinforcement Learning (RL). To address these challenges, we aim to develop a versatile diffusion planner that can leverage large-scale inferior data that contains task-agnostic sub-optimal trajectories, with the ability to fast adapt to specific tasks. In this paper, we propose \textbf{SODP}, a two-stage framework that leverages \textbf{S}ub-\textbf{O}ptimal data to learn a \textbf{D}iffusion \textbf{P}lanner, which is generalizable for various downstream tasks. Specifically, in the pre-training stage, we train a foundation diffusion planner that extracts general planning capabilities by modeling the versatile distribution of multi-task trajectories, which can be sub-optimal and has wide data coverage. Then for downstream tasks, we adopt RL-based fine-tuning with task-specific rewards to fast refine the diffusion planner, which aims to generate action sequences with higher task-specific returns. Experimental results from multi-task domains including Meta-World and Adroit demonstrate that SODP outperforms state-of-the-art methods with only a small amount of data for reward-guided fine-tuning.

Learning a world model for model-free Reinforcement Learning (RL) agents can significantly improve the sample efficiency by learning policies in imagination. However, building a world model for Multi-Agent RL (MARL) can be particularly challenging due to the scalability issue in a centralized architecture arising from a large number of agents, and also the non-stationarity issue in a decentralized architecture stemming from the inter-dependency among agents. To address both challenges, we propose a novel world model for MARL that learns decentralized local dynamics for scalability, combined with a centralized representation aggregation from all agents. We cast the dynamics learning as an auto-regressive sequence modeling problem over discrete tokens by leveraging the expressive Transformer architecture, in order to model complex local dynamics across different agents and provide accurate and consistent long-term imaginations. As the first pioneering Transformer-based world model for multi-agent systems, we introduce a Perceiver Transformer as an effective solution to enable centralized representation aggregation within this context. Results on Starcraft Multi-Agent Challenge (SMAC) show that it outperforms strong model-free approaches and existing model-based methods in both sample efficiency and overall performance.

Acquiring a multi-task imitation policy in 3D manipulation poses challenges in terms of scene understanding and action prediction. Current methods employ both 3D representation and multi-view 2D representation to predict the poses of the robot's end-effector. However, they still require a considerable amount of high-quality robot trajectories, and suffer from limited generalization in unseen tasks and inefficient execution in long-horizon reasoning. In this paper, we propose SAM-E, a novel architecture for robot manipulation by leveraging a vision-foundation model for generalizable scene understanding and sequence imitation for long-term action reasoning. Specifically, we adopt Segment Anything (SAM) pre-trained on a huge number of images and promptable masks as the foundation model for extracting task-relevant features, and employ parameter-efficient fine-tuning on robot data for a better understanding of embodied scenarios. To address long-horizon reasoning, we develop a novel multi-channel heatmap that enables the prediction of the action sequence in a single pass, notably enhancing execution efficiency. Experimental results from various instruction-following tasks demonstrate that SAM-E achieves superior performance with higher execution efficiency compared to the baselines, and also significantly improves generalization in few-shot adaptation to new tasks.