Zhengbang Zhu

Zhengbang Zhu, Hanye Zhao, Haoran He et al.

Diffusion models surpass previous generative models in sample quality and training stability. Recent works have shown the advantages of diffusion models in improving reinforcement learning (RL) solutions. This survey aims to provide an overview of this emerging field and hopes to inspire new avenues of research. First, we examine several challenges encountered by RL algorithms. Then, we present a taxonomy of existing methods based on the roles of diffusion models in RL and explore how the preceding challenges are addressed. We further outline successful applications of diffusion models in various RL-related tasks. Finally, we conclude the survey and offer insights into future research directions. We are actively maintaining a GitHub repository for papers and other related resources in utilizing diffusion models in RL: https://github.com/apexrl/Diff4RLSurvey.

Hanye Zhao, Xiaoshen Han, Zhengbang Zhu et al.

With the great success of diffusion models (DMs) in generating realistic synthetic vision data, many researchers have investigated their potential in decision-making and control. Most of these works utilized DMs to sample directly from the trajectory space, where DMs can be viewed as a combination of dynamics models and policies. In this work, we explore how to decouple DMs' ability as dynamics models in fully offline settings, allowing the learning policy to roll out trajectories. As DMs learn the data distribution from the dataset, their intrinsic policy is actually the behavior policy induced from the dataset, which results in a mismatch between the behavior policy and the learning policy. We propose Dynamics Diffusion, short as DyDiff, which can inject information from the learning policy to DMs iteratively. DyDiff ensures long-horizon rollout accuracy while maintaining policy consistency and can be easily deployed on model-free algorithms. We provide theoretical analysis to show the advantage of DMs on long-horizon rollout over models and demonstrate the effectiveness of DyDiff in the context of offline reinforcement learning, where the rollout dataset is provided but no online environment for interaction.

Minghuan Liu, Zhengbang Zhu, Yuzheng Zhuang et al.

Recent progress in state-only imitation learning extends the scope of applicability of imitation learning to real-world settings by relieving the need for observing expert actions. However, existing solutions only learn to extract a state-to-action mapping policy from the data, without considering how the expert plans to the target. This hinders the ability to leverage demonstrations and limits the flexibility of the policy. In this paper, we introduce Decoupled Policy Optimization (DePO), which explicitly decouples the policy as a high-level state planner and an inverse dynamics model. With embedded decoupled policy gradient and generative adversarial training, DePO enables knowledge transfer to different action spaces or state transition dynamics, and can generalize the planner to out-of-demonstration state regions. Our in-depth experimental analysis shows the effectiveness of DePO on learning a generalized target state planner while achieving the best imitation performance. We demonstrate the appealing usage of DePO for transferring across different tasks by pre-training, and the potential for co-training agents with various skills.

Zhengbang Zhu, Shenyu Zhang, Yuzheng Zhuang et al.

High-quality traffic flow generation is the core module in building simulators for autonomous driving. However, the majority of available simulators are incapable of replicating traffic patterns that accurately reflect the various features of real-world data while also simulating human-like reactive responses to the tested autopilot driving strategies. Taking one step forward to addressing such a problem, we propose Realistic Interactive TrAffic flow (RITA) as an integrated component of existing driving simulators to provide high-quality traffic flow for the evaluation and optimization of the tested driving strategies. RITA is developed with consideration of three key features, i.e., fidelity, diversity, and controllability, and consists of two core modules called RITABackend and RITAKit. RITABackend is built to support vehicle-wise control and provide traffic generation models from real-world datasets, while RITAKit is developed with easy-to-use interfaces for controllable traffic generation via RITABackend. We demonstrate RITA's capacity to create diversified and high-fidelity traffic simulations in several highly interactive highway scenarios. The experimental findings demonstrate that our produced RITA traffic flows exhibit all three key features, hence enhancing the completeness of driving strategy evaluation. Moreover, we showcase the possibility for further improvement of baseline strategies through online fine-tuning with RITA traffic flows.

Zhengbang Zhu, Rongjun Qin, Junjie Huang et al.

Recommender systems are expected to be assistants that help human users find relevant information automatically without explicit queries. As recommender systems evolve, increasingly sophisticated learning techniques are applied and have achieved better performance in terms of user engagement metrics such as clicks and browsing time. The increase in the measured performance, however, can have two possible attributions: a better understanding of user preferences, and a more proactive ability to utilize human bounded rationality to seduce user over-consumption. A natural following question is whether current recommendation algorithms are manipulating user preferences. If so, can we measure the manipulation level? In this paper, we present a general framework for benchmarking the degree of manipulations of recommendation algorithms, in both slate recommendation and sequential recommendation scenarios. The framework consists of four stages, initial preference calculation, training data collection, algorithm training and interaction, and metrics calculation that involves two proposed metrics. We benchmark some representative recommendation algorithms in both synthetic and real-world datasets under the proposed framework. We have observed that a high online click-through rate does not necessarily mean a better understanding of user initial preference, but ends in prompting users to choose more documents they initially did not favor. Moreover, we find that the training data have notable impacts on the manipulation degrees, and algorithms with more powerful modeling abilities are more sensitive to such impacts. The experiments also verified the usefulness of the proposed metrics for measuring the degree of manipulations. We advocate that future recommendation algorithm studies should be treated as an optimization problem with constrained user preference manipulations.

Minghuan Liu, Zhengbang Zhu, Menghui Zhu et al.

In reinforcement learning applications like robotics, agents usually need to deal with various input/output features when specified with different state/action spaces by their developers or physical restrictions. This indicates unnecessary re-training from scratch and considerable sample inefficiency, especially when agents follow similar solution steps to achieve tasks. In this paper, we aim to transfer similar high-level goal-transition knowledge to alleviate the challenge. Specifically, we propose PILoT, i.e., Planning Immediate Landmarks of Targets. PILoT utilizes the universal decoupled policy optimization to learn a goal-conditioned state planner; then, distills a goal-planner to plan immediate landmarks in a model-free style that can be shared among different agents. In our experiments, we show the power of PILoT on various transferring challenges, including few-shot transferring across action spaces and dynamics, from low-dimensional vector states to image inputs, from simple robot to complicated morphology; and we also illustrate a zero-shot transfer solution from a simple 2D navigation task to the harder Ant-Maze task.

Yue Su, Sijin Chen, Haixin Shi et al.

Leveraging future observation modeling to facilitate action generation presents a promising avenue for enhancing the capabilities of Vision-Language-Action (VLA) models. However, existing approaches struggle to strike a balance between maintaining efficient, predictable future representations and preserving sufficient fine-grained information to guide precise action generation. To address this limitation, we propose WoG (World Guidance), a framework that maps future observations into compact conditions by injecting them into the action inference pipeline. The VLA is then trained to simultaneously predict these compressed conditions alongside future actions, thereby achieving effective world modeling within the condition space for action inference. We demonstrate that modeling and predicting this condition space not only facilitates fine-grained action generation but also exhibits superior generalization capabilities. Moreover, it learns effectively from substantial human manipulation videos. Extensive experiments across both simulation and real-world environments validate that our method significantly outperforms existing methods based on future prediction. Project page is available at: https://selen-suyue.github.io/WoGNet/

Jiaguo Tian, Zhengbang Zhu, Shenyu Zhang et al.

The generation of realistic and diverse traffic scenarios in simulation is essential for developing and evaluating autonomous driving systems. However, most simulation frameworks rely on rule-based or simplified models for scene generation, which lack the fidelity and diversity needed to represent real-world driving. While recent advances in generative modeling produce more realistic and context-aware traffic interactions, they often overlook how social preferences influence driving behavior. SocialDriveGen addresses this gap through a hierarchical framework that integrates semantic reasoning and social preference modeling with generative trajectory synthesis. By modeling egoism and altruism as complementary social dimensions, our framework enables controllable diversity in driver personalities and interaction styles. Experiments on the Argoverse 2 dataset show that SocialDriveGen generates diverse, high-fidelity traffic scenarios spanning cooperative to adversarial behaviors, significantly enhancing policy robustness and generalization to rare or high-risk situations.

Yixiang Shan, Zhengbang Zhu, Ting Long et al.

The performance of offline reinforcement learning (RL) is sensitive to the proportion of high-return trajectories in the offline dataset. However, in many simulation environments and real-world scenarios, there are large ratios of low-return trajectories rather than high-return trajectories, which makes learning an efficient policy challenging. In this paper, we propose a method called Contrastive Diffuser (CDiffuser) to make full use of low-return trajectories and improve the performance of offline RL algorithms. Specifically, CDiffuser groups the states of trajectories in the offline dataset into high-return states and low-return states and treats them as positive and negative samples correspondingly. Then, it designs a contrastive mechanism to pull the trajectory of an agent toward high-return states and push them away from low-return states. Through the contrast mechanism, trajectories with low returns can serve as negative examples for policy learning, guiding the agent to avoid areas associated with low returns and achieve better performance. Experiments on 14 commonly used D4RL benchmarks demonstrate the effectiveness of our proposed method. Our code is publicly available at \url{https://anonymous.4open.science/r/CDiffuser}.

Zhengbang Zhu, Minghuan Liu, Liyuan Mao et al.

Offline reinforcement learning (RL) aims to learn policies from pre-existing datasets without further interactions, making it a challenging task. Q-learning algorithms struggle with extrapolation errors in offline settings, while supervised learning methods are constrained by model expressiveness. Recently, diffusion models (DMs) have shown promise in overcoming these limitations in single-agent learning, but their application in multi-agent scenarios remains unclear. Generating trajectories for each agent with independent DMs may impede coordination, while concatenating all agents' information can lead to low sample efficiency. Accordingly, we propose MADiff, which is realized with an attention-based diffusion model to model the complex coordination among behaviors of multiple agents. To our knowledge, MADiff is the first diffusion-based multi-agent learning framework, functioning as both a decentralized policy and a centralized controller. During decentralized executions, MADiff simultaneously performs teammate modeling, and the centralized controller can also be applied in multi-agent trajectory predictions. Our experiments demonstrate that MADiff outperforms baseline algorithms across various multi-agent learning tasks, highlighting its effectiveness in modeling complex multi-agent interactions. Our code is available at https://github.com/zbzhu99/madiff.

Ming Zhou, Jun Luo, Julian Villella et al.

Multi-agent interaction is a fundamental aspect of autonomous driving in the real world. Despite more than a decade of research and development, the problem of how to competently interact with diverse road users in diverse scenarios remains largely unsolved. Learning methods have much to offer towards solving this problem. But they require a realistic multi-agent simulator that generates diverse and competent driving interactions. To meet this need, we develop a dedicated simulation platform called SMARTS (Scalable Multi-Agent RL Training School). SMARTS supports the training, accumulation, and use of diverse behavior models of road users. These are in turn used to create increasingly more realistic and diverse interactions that enable deeper and broader research on multi-agent interaction. In this paper, we describe the design goals of SMARTS, explain its basic architecture and its key features, and illustrate its use through concrete multi-agent experiments on interactive scenarios. We open-source the SMARTS platform and the associated benchmark tasks and evaluation metrics to encourage and empower research on multi-agent learning for autonomous driving. Our code is available at https://github.com/huawei-noah/SMARTS.

Guanghe Li, Yixiang Shan, Zhengbang Zhu et al.

In offline reinforcement learning (RL), the performance of the learned policy highly depends on the quality of offline datasets. However, in many cases, the offline dataset contains very limited optimal trajectories, which poses a challenge for offline RL algorithms as agents must acquire the ability to transit to high-reward regions. To address this issue, we introduce Diffusion-based Trajectory Stitching (DiffStitch), a novel diffusion-based data augmentation pipeline that systematically generates stitching transitions between trajectories. DiffStitch effectively connects low-reward trajectories with high-reward trajectories, forming globally optimal trajectories to address the challenges faced by offline RL algorithms. Empirical experiments conducted on D4RL datasets demonstrate the effectiveness of DiffStitch across RL methodologies. Notably, DiffStitch demonstrates substantial enhancements in the performance of one-step methods (IQL), imitation learning methods (TD3+BC), and trajectory optimization methods (DT).

Shenyu Zhang, Jiaguo Tian, Zhengbang Zhu et al.

Microscopic traffic simulation has become an important tool for autonomous driving training and testing. Although recent data-driven approaches advance realistic behavior generation, their learning still relies primarily on a single real-world dataset, which limits their diversity and thereby hinders downstream algorithm optimization. In this paper, we propose DriveGen, a novel traffic simulation framework with large models for more diverse traffic generation that supports further customized designs. DriveGen consists of two internal stages: the initialization stage uses large language model and retrieval technique to generate map and vehicle assets; the rollout stage outputs trajectories with selected waypoint goals from visual language model and a specific designed diffusion planner. Through this two-staged process, DriveGen fully utilizes large models' high-level cognition and reasoning of driving behavior, obtaining greater diversity beyond datasets while maintaining high realism. To support effective downstream optimization, we additionally develop DriveGen-CS, an automatic corner case generation pipeline that uses failures of the driving algorithm as additional prompt knowledge for large models without the need for retraining or fine-tuning. Experiments show that our generated scenarios and corner cases have a superior performance compared to state-of-the-art baselines. Downstream experiments further verify that the synthesized traffic of DriveGen provides better optimization of the performance of typical driving algorithms, demonstrating the effectiveness of our framework.

Minghuan Liu, Zhengbang Zhu, Xiaoshen Han et al.

Modern robotic manipulation primarily relies on visual observations in a 2D color space for skill learning but suffers from poor generalization. In contrast, humans, living in a 3D world, depend more on physical properties-such as distance, size, and shape-than on texture when interacting with objects. Since such 3D geometric information can be acquired from widely available depth cameras, it appears feasible to endow robots with similar perceptual capabilities. Our pilot study found that using depth cameras for manipulation is challenging, primarily due to their limited accuracy and susceptibility to various types of noise. In this work, we propose Camera Depth Models (CDMs) as a simple plugin on daily-use depth cameras, which take RGB images and raw depth signals as input and output denoised, accurate metric depth. To achieve this, we develop a neural data engine that generates high-quality paired data from simulation by modeling a depth camera's noise pattern. Our results show that CDMs achieve nearly simulation-level accuracy in depth prediction, effectively bridging the sim-to-real gap for manipulation tasks. Notably, our experiments demonstrate, for the first time, that a policy trained on raw simulated depth, without the need for adding noise or real-world fine-tuning, generalizes seamlessly to real-world robots on two challenging long-horizon tasks involving articulated, reflective, and slender objects, with little to no performance degradation. We hope our findings will inspire future research in utilizing simulation data and 3D information in general robot policies.

Jingxiao Chen, Xinyao Li, Jiahang Cao et al.

Humanoid robots have shown success in locomotion and manipulation. Despite these basic abilities, humanoids are still required to quickly understand human instructions and react based on human interaction signals to become valuable assistants in human daily life. Unfortunately, most existing works only focus on multi-stage interactions, treating each task separately, and neglecting real-time feedback. In this work, we aim to empower humanoid robots with real-time reaction abilities to achieve various tasks, allowing human to interrupt robots at any time, and making robots respond to humans immediately. To support such abilities, we propose a general humanoid-human-object interaction framework, named RHINO, i.e., Real-time Humanoid-human Interaction and Object manipulation. RHINO provides a unified view of reactive motion, instruction-based manipulation, and safety concerns, over multiple human signal modalities, such as languages, images, and motions. RHINO is a hierarchical learning framework, enabling humanoids to learn reaction skills from human-human-object demonstrations and teleoperation data. In particular, it decouples the interaction process into two levels: 1) a high-level planner inferring human intentions from real-time human behaviors; and 2) a low-level controller achieving reactive motion behaviors and object manipulation skills based on the predicted intentions. We evaluate the proposed framework on a real humanoid robot and demonstrate its effectiveness, flexibility, and safety in various scenarios.

Zhengshen Zhang, Hao Li, Yalun Dai et al.

Existing vision-language-action (VLA) models act in 3D real-world but are typically built on 2D encoders, leaving a spatial reasoning gap that limits generalization and adaptability. Recent 3D integration techniques for VLAs either require specialized sensors and transfer poorly across modalities, or inject weak cues that lack geometry and degrade vision-language alignment. In this work, we introduce FALCON (From Spatial to Action), a novel paradigm that injects rich 3D spatial tokens into the action head. FALCON leverages spatial foundation models to deliver strong geometric priors from RGB alone, and includes an Embodied Spatial Model that can optionally fuse depth, or pose for higher fidelity when available, without retraining or architectural changes. To preserve language reasoning, spatial tokens are consumed by a Spatial-Enhanced Action Head rather than being concatenated into the vision-language backbone. These designs enable FALCON to address limitations in spatial representation, modality transferability, and alignment. In comprehensive evaluations across three simulation benchmarks and eleven real-world tasks, our proposed FALCON achieves state-of-the-art performance, consistently surpasses competitive baselines, and remains robust under clutter, spatial-prompt conditioning, and variations in object scale and height.

Lu Guo, Yixiang Shan, Zhengbang Zhu et al.

Offline reinforcement learning (RL) enables agents to learn policies from fixed datasets, avoiding costly or unsafe environment interactions. However, its effectiveness is often limited by dataset sparsity and the lack of transition overlap between suboptimal and expert trajectories, which makes long-horizon planning particularly challenging. Prior solutions based on synthetic data augmentation or trajectory stitching often fail to generalize to novel states and rely on heuristic stitching points. To address these challenges, we propose Retrieval High-quAlity Demonstrations (RAD) for decision-making, which combines non-parametric retrieval with diffusion-based generative modeling. RAD dynamically retrieves high-return states from the offline dataset as target states based on state similarity and return estimation, and plans toward them using a condition-guided diffusion model. Such retrieval-guided generation enables flexible trajectory stitching and improves generalization when encountered with underrepresented or out-of-distribution states. Extensive experiments confirm that RAD achieves competitive or superior performance compared to baselines across diverse benchmarks, validating its effectiveness.