Zifeng Zhuang

Zifeng Zhuang, Kun Lei, Jinxin Liu et al.

Offline reinforcement learning (RL) is a challenging setting where existing off-policy actor-critic methods perform poorly due to the overestimation of out-of-distribution state-action pairs. Thus, various additional augmentations are proposed to keep the learned policy close to the offline dataset (or the behavior policy). In this work, starting from the analysis of offline monotonic policy improvement, we get a surprising finding that some online on-policy algorithms are naturally able to solve offline RL. Specifically, the inherent conservatism of these on-policy algorithms is exactly what the offline RL method needs to overcome the overestimation. Based on this, we propose Behavior Proximal Policy Optimization (BPPO), which solves offline RL without any extra constraint or regularization introduced compared to PPO. Extensive experiments on the D4RL benchmark indicate this extremely succinct method outperforms state-of-the-art offline RL algorithms. Our implementation is available at https://github.com/Dragon-Zhuang/BPPO.

Runze Li, Hongyin Zhang, Junxi Jin et al.

Vision-Language-Action (VLA) models have emerged as a promising paradigm for building embodied agents that ground perception and language into action. However, most existing approaches rely on direct action prediction, lacking the ability to reason over long-horizon trajectories and evaluate their consequences, which limits performance in complex decision-making tasks. In this work, we introduce World-Value-Action (WAV) model, a unified framework that enables implicit planning in VLA systems. Rather than performing explicit trajectory optimization, WAV model learn a structured latent representation of future trajectories conditioned on visual observations and language instructions. A learned world model predicts future states, while a trajectory value function evaluates their long-horizon utility. Action generation is then formulated as inference in this latent space, where the model progressively concentrates probability mass on high-value and dynamically feasible trajectories. We provide a theoretical perspective showing that planning directly in action space suffers from an exponential decay in the probability of feasible trajectories as the horizon increases. In contrast, latent-space inference reshapes the search distribution toward feasible regions, enabling efficient long-horizon decision making. Extensive simulations and real-world experiments demonstrate that the WAV model consistently outperforms state-of-the-art methods, achieving significant improvements in task success rate, generalization ability, and robustness, especially in long-horizon and compositional scenarios. Code is available at https://github.com/Win-commit/WAV.

Jinxin Liu, Li He, Yachen Kang et al.

In this paper, we present \textbf{C}ont\textbf{E}xtual \textbf{I}mitation \textbf{L}earning~(CEIL), a general and broadly applicable algorithm for imitation learning (IL). Inspired by the formulation of hindsight information matching, we derive CEIL by explicitly learning a hindsight embedding function together with a contextual policy using the hindsight embeddings. To achieve the expert matching objective for IL, we advocate for optimizing a contextual variable such that it biases the contextual policy towards mimicking expert behaviors. Beyond the typical learning from demonstrations (LfD) setting, CEIL is a generalist that can be effectively applied to multiple settings including: 1)~learning from observations (LfO), 2)~offline IL, 3)~cross-domain IL (mismatched experts), and 4) one-shot IL settings. Empirically, we evaluate CEIL on the popular MuJoCo tasks (online) and the D4RL dataset (offline). Compared to prior state-of-the-art baselines, we show that CEIL is more sample-efficient in most online IL tasks and achieves better or competitive performances in offline tasks.

Jinxin Liu, Ziqi Zhang, Zhenyu Wei et al.

Offline reinforcement learning (RL) aims to learn a policy using only pre-collected and fixed data. Although avoiding the time-consuming online interactions in RL, it poses challenges for out-of-distribution (OOD) state actions and often suffers from data inefficiency for training. Despite many efforts being devoted to addressing OOD state actions, the latter (data inefficiency) receives little attention in offline RL. To address this, this paper proposes the cross-domain offline RL, which assumes offline data incorporate additional source-domain data from varying transition dynamics (environments), and expects it to contribute to the offline data efficiency. To do so, we identify a new challenge of OOD transition dynamics, beyond the common OOD state actions issue, when utilizing cross-domain offline data. Then, we propose our method BOSA, which employs two support-constrained objectives to address the above OOD issues. Through extensive experiments in the cross-domain offline RL setting, we demonstrate BOSA can greatly improve offline data efficiency: using only 10\% of the target data, BOSA could achieve {74.4\%} of the SOTA offline RL performance that uses 100\% of the target data. Additionally, we also show BOSA can be effortlessly plugged into model-based offline RL and noising data augmentation techniques (used for generating source-domain data), which naturally avoids the potential dynamics mismatch between target-domain data and newly generated source-domain data.

Jinxin Liu, Hongyin Zhang, Zifeng Zhuang et al.

In this work, we decouple the iterative bi-level offline RL (value estimation and policy extraction) from the offline training phase, forming a non-iterative bi-level paradigm and avoiding the iterative error propagation over two levels. Specifically, this non-iterative paradigm allows us to conduct inner-level optimization (value estimation) in training, while performing outer-level optimization (policy extraction) in testing. Naturally, such a paradigm raises three core questions that are not fully answered by prior non-iterative offline RL counterparts like reward-conditioned policy: (q1) What information should we transfer from the inner-level to the outer-level? (q2) What should we pay attention to when exploiting the transferred information for safe/confident outer-level optimization? (q3) What are the benefits of concurrently conducting outer-level optimization during testing? Motivated by model-based optimization (MBO), we propose DROP (design from policies), which fully answers the above questions. Specifically, in the inner-level, DROP decomposes offline data into multiple subsets, and learns an MBO score model (a1). To keep safe exploitation to the score model in the outer-level, we explicitly learn a behavior embedding and introduce a conservative regularization (a2). During testing, we show that DROP permits deployment adaptation, enabling an adaptive inference across states (a3). Empirically, we evaluate DROP on various tasks, showing that DROP gains comparable or better performance compared to prior methods.

Minghui Lin, Pengxiang Ding, Shu Wang et al.

Vision-Language-Action (VLA) models have recently enabled robotic manipulation by grounding visual and linguistic cues into actions. However, most VLAs assume the Markov property, relying only on the current observation and thus suffering from temporal myopia that degrades long-horizon coherence. In this work, we view motion as a more compact and informative representation of temporal context and world dynamics, capturing inter-state changes while filtering static pixel-level noise. From this perspective, HiF-VLA equips a motion-centric world model for the VLA, enabling agents to reason about temporal dynamics for future evolution during action generation. Building on this idea, we propose HiF-VLA (Hindsight, Insight, and Foresight for VLAs), a unified framework that leverages motion for bidirectional temporal reasoning. HiF-VLA encodes past dynamics through hindsight priors, anticipates future motion via foresight reasoning, and integrates both through a hindsight-modulated joint expert to enable a ''think-while-acting'' paradigm for long-horizon manipulation. As a result, HiF-VLA surpasses strong baselines on LIBERO-Long and CALVIN ABC-D benchmarks, while incurring negligible additional inference latency. Furthermore, HiF-VLA achieves substantial improvements in real-world long-horizon manipulation tasks, demonstrating its broad effectiveness in practical robotic settings.

Min Zhang, Zifeng Zhuang, Zhitao Wang et al.

Gradient-based meta-learning (GBML) algorithms are able to fast adapt to new tasks by transferring the learned meta-knowledge, while assuming that all tasks come from the same distribution (in-distribution, ID). However, in the real world, they often suffer from an out-of-distribution (OOD) generalization problem, where tasks come from different distributions. OOD exacerbates inconsistencies in magnitudes and directions of task gradients, which brings challenges for GBML to optimize the meta-knowledge by minimizing the sum of task gradients in each minibatch. To address this problem, we propose RotoGBML, a novel approach to homogenize OOD task gradients. RotoGBML uses reweighted vectors to dynamically balance diverse magnitudes to a common scale and uses rotation matrixes to rotate conflicting directions close to each other. To reduce overhead, we homogenize gradients with the features rather than the network parameters. On this basis, to avoid the intervention of non-causal features (e.g., backgrounds), we also propose an invariant self-information (ISI) module to extract invariant causal features (e.g., the outlines of objects). Finally, task gradients are homogenized based on these invariant causal features. Experiments show that RotoGBML outperforms other state-of-the-art methods on various few-shot image classification benchmarks.

Yang Liu, Pengxiang Ding, Tengyue Jiang et al.

Vision-Language-Action (VLA) models aim to control robots for manipulation from visual observations and natural-language instructions. However, existing hierarchical and autoregressive paradigms often introduce architectural overhead, suffer from temporal inconsistency and long-horizon error accumulation, and lack a mechanism to capture environment dynamics without extra modules. To this end, we present MMaDA-VLA, a fully native pre-trained large diffusion VLA model that unifies multi-modal understanding and generation in a single framework. Our key idea is a native discrete diffusion formulation that embeds language, images, and continuous robot controls into one discrete token space and trains a single backbone with masked token denoising to jointly generate a future goal observation and an action chunk in parallel. Iterative denoising enables global, order-free refinement, improving long-horizon consistency while grounding actions in predicted future visual outcomes without auxiliary world models. Experiments across simulation benchmarks and real-world tasks show state-of-the-art performance, achieving 98.0% average success on LIBERO and 4.78 average length on CALVIN.

Yachen Kang, Li He, Jinxin Liu et al.

Preference-based reinforcement learning (PbRL) promises to learn a complex reward function with binary human preference. However, such human-in-the-loop formulation requires considerable human effort to assign preference labels to segment pairs, hindering its large-scale applications. Recent approache has tried to reuse unlabeled segments, which implicitly elucidates the distribution of segments and thereby alleviates the human effort. And consistency regularization is further considered to improve the performance of semi-supervised learning. However, we notice that, unlike general classification tasks, in PbRL there exits a unique phenomenon that we defined as similarity trap in this paper. Intuitively, human can have diametrically opposite preferredness for similar segment pairs, but such similarity may trap consistency regularization fail in PbRL. Due to the existence of similarity trap, such consistency regularization improperly enhances the consistency possiblity of the model's predictions between segment pairs, and thus reduces the confidence in reward learning, since the augmented distribution does not match with the original one in PbRL. To overcome such issue, we present a self-training method along with our proposed peer regularization, which penalizes the reward model memorizing uninformative labels and acquires confident predictions. Empirically, we demonstrate that our approach is capable of learning well a variety of locomotion and robotic manipulation behaviors using different semi-supervised alternatives and peer regularization.

Zifeng Zhuang, Dengyun Peng, Jinxin Liu et al.

As a data-driven paradigm, offline reinforcement learning (RL) has been formulated as sequence modeling that conditions on the hindsight information including returns, goal or future trajectory. Although promising, this supervised paradigm overlooks the core objective of RL that maximizes the return. This overlook directly leads to the lack of trajectory stitching capability that affects the sequence model learning from sub-optimal data. In this work, we introduce the concept of max-return sequence modeling which integrates the goal of maximizing returns into existing sequence models. We propose Reinforced Transformer (Reinformer), indicating the sequence model is reinforced by the RL objective. Reinformer additionally incorporates the objective of maximizing returns in the training phase, aiming to predict the maximum future return within the distribution. During inference, this in-distribution maximum return will guide the selection of optimal actions. Empirically, Reinformer is competitive with classical RL methods on the D4RL benchmark and outperforms state-of-the-art sequence model particularly in trajectory stitching ability. Code is public at https://github.com/Dragon-Zhuang/Reinformer.

Hongyin Zhang, Diyuan Shi, Zifeng Zhuang et al.

Developing robotic intelligent systems that can adapt quickly to unseen wild situations is one of the critical challenges in pursuing autonomous robotics. Although some impressive progress has been made in walking stability and skill learning in the field of legged robots, their ability for fast adaptation is still inferior to that of animals in nature. Animals are born with a massive set of skills needed to survive, and can quickly acquire new ones, by composing fundamental skills with limited experience. Inspired by this, we propose a novel framework, named Robot Skill Graph (RSG) for organizing a massive set of fundamental skills of robots and dexterously reusing them for fast adaptation. Bearing a structure similar to the Knowledge Graph (KG), RSG is composed of massive dynamic behavioral skills instead of static knowledge in KG and enables discovering implicit relations that exist in between the learning context and acquired skills of robots, serving as a starting point for understanding subtle patterns existing in robots' skill learning. Extensive experimental results demonstrate that RSG can provide reliable skill inference upon new tasks and environments, and enable quadruped robots to adapt to new scenarios and quickly learn new skills.

Ziqi Zhang, Xiao Xiong, Zifeng Zhuang et al.

Studying how to fine-tune offline reinforcement learning (RL) pre-trained policy is profoundly significant for enhancing the sample efficiency of RL algorithms. However, directly fine-tuning pre-trained policies often results in sub-optimal performance. This is primarily due to the distribution shift between offline pre-training and online fine-tuning stages. Specifically, the distribution shift limits the acquisition of effective online samples, ultimately impacting the online fine-tuning performance. In order to narrow down the distribution shift between offline and online stages, we proposed Q conditioned state entropy (QCSE) as intrinsic reward. Specifically, QCSE maximizes the state entropy of all samples individually, considering their respective Q values. This approach encourages exploration of low-frequency samples while penalizing high-frequency ones, and implicitly achieves State Marginal Matching (SMM), thereby ensuring optimal performance, solving the asymptotic sub-optimality of constraint-based approaches. Additionally, QCSE can seamlessly integrate into various RL algorithms, enhancing online fine-tuning performance. To validate our claim, we conduct extensive experiments, and observe significant improvements with QCSE (about 13% for CQL and 8% for Cal-QL). Furthermore, we extended experimental tests to other algorithms, affirming the generality of QCSE.

Diyuan Shi, Yiqi Tang, Zifeng Zhuang et al.

Deep Reinforcement Learning (DRL) has experienced significant advancements in recent years and has been widely used in many fields. In DRL-based robotic policy learning, however, current de facto policy parameterization is still multivariate Gaussian (with diagonal covariance matrix), which lacks the ability to model multi-modal distribution. In this work, we explore the adoption of a modern network architecture, i.e. Normalizing Flow (NF) as the policy parameterization for its ability of multi-modal modeling, closed form of log probability and low computation and memory overhead. However, naively training NF in online Reinforcement Learning (RL) usually leads to training instability. We provide a detailed analysis for this phenomenon and successfully address it via simple but effective technique. With extensive experiments in multiple simulation environments, we show our method, NFPO could obtain robust and strong performance in widely used robotic learning tasks and successfully transfer into real-world robots.

Ziqi Zhang, Jingzehua Xu, Zifeng Zhuang et al.

Proximal Policy Optimization (PPO) has been broadly applied to robotics learning, showcasing stable training performance. However, the fixed clipping bound setting may limit the performance of PPO. Specifically, there is no theoretical proof that the optimal clipping bound remains consistent throughout the entire training process. Meanwhile, previous researches suggest that a fixed clipping bound restricts the policy's ability to explore. Therefore, many past studies have aimed to dynamically adjust the PPO clipping bound to enhance PPO's performance. However, the objective of these approaches are not directly aligned with the objective of reinforcement learning (RL) tasks, which is to maximize the cumulative Return. Unlike previous clipping approaches, we propose a bi-level proximal policy optimization objective that can dynamically adjust the clipping bound to better reflect the preference (maximizing Return) of these RL tasks. Based on this bi-level proximal policy optimization paradigm, we introduce a new algorithm named Preference based Proximal Policy Optimization (Pb-PPO). Pb-PPO utilizes a multi-armed bandit approach to refelect RL preference, recommending the clipping bound for PPO that can maximizes the current Return. Therefore, Pb-PPO results in greater stability and improved performance compared to PPO with a fixed clipping bound. We test Pb-PPO on locomotion benchmarks across multiple environments, including Gym-Mujoco and legged-gym. Additionally, we validate Pb-PPO on customized navigation tasks. Meanwhile, we conducted comparisons with PPO using various fixed clipping bounds and various of clipping approaches. The experimental results indicate that Pb-PPO demonstrates superior training performance compared to PPO and its variants. Our codebase has been released at : https://github.com/stevezhangzA/pb_ppo

Huikang Su, Dengyun Peng, Zifeng Zhuang et al.

Offline safe reinforcement learning aims to learn policies that satisfy predefined safety constraints from static datasets. Existing sequence-model-based methods condition action generation on symmetric input tokens for return-to-go and cost-to-go, neglecting their intrinsic asymmetry: return-to-go (RTG) serves as a flexible performance target, while cost-to-go (CTG) should represent a rigid safety boundary. This symmetric conditioning leads to unreliable constraint satisfaction, especially when encountering out-of-distribution cost trajectories. To address this, we propose Boundary-to-Region (B2R), a framework that enables asymmetric conditioning through cost signal realignment . B2R redefines CTG as a boundary constraint under a fixed safety budget, unifying the cost distribution of all feasible trajectories while preserving reward structures. Combined with rotary positional embeddings , it enhances exploration within the safe region. Experimental results show that B2R satisfies safety constraints in 35 out of 38 safety-critical tasks while achieving superior reward performance over baseline methods. This work highlights the limitations of symmetric token conditioning and establishes a new theoretical and practical approach for applying sequence models to safe RL. Our code is available at https://github.com/HuikangSu/B2R.

Ziqi Zhang, Jingzehua Xu, Jinxin Liu et al.

Offline reinforcement learning (RL) algorithms can learn better decision-making compared to behavior policies by stitching the suboptimal trajectories to derive more optimal ones. Meanwhile, Decision Transformer (DT) abstracts the RL as sequence modeling, showcasing competitive performance on offline RL benchmarks. However, recent studies demonstrate that DT lacks of stitching capacity, thus exploiting stitching capability for DT is vital to further improve its performance. In order to endow stitching capability to DT, we abstract trajectory stitching as expert matching and introduce our approach, ContextFormer, which integrates contextual information-based imitation learning (IL) and sequence modeling to stitch sub-optimal trajectory fragments by emulating the representations of a limited number of expert trajectories. To validate our approach, we conduct experiments from two perspectives: 1) We conduct extensive experiments on D4RL benchmarks under the settings of IL, and experimental results demonstrate ContextFormer can achieve competitive performance in multiple IL settings. 2) More importantly, we conduct a comparison of ContextFormer with various competitive DT variants using identical training datasets. The experimental results unveiled ContextFormer's superiority, as it outperformed all other variants, showcasing its remarkable performance.

Jinxin Liu, Xinghong Guo, Zifeng Zhuang et al.

In this paper, we propose a novel approach called DIffusion-guided DIversity (DIDI) for offline behavioral generation. The goal of DIDI is to learn a diverse set of skills from a mixture of label-free offline data. We achieve this by leveraging diffusion probabilistic models as priors to guide the learning process and regularize the policy. By optimizing a joint objective that incorporates diversity and diffusion-guided regularization, we encourage the emergence of diverse behaviors while maintaining the similarity to the offline data. Experimental results in four decision-making domains (Push, Kitchen, Humanoid, and D4RL tasks) show that DIDI is effective in discovering diverse and discriminative skills. We also introduce skill stitching and skill interpolation, which highlight the generalist nature of the learned skill space. Further, by incorporating an extrinsic reward function, DIDI enables reward-guided behavior generation, facilitating the learning of diverse and optimal behaviors from sub-optimal data.

Wei Xiao, Jiacheng Liu, Zifeng Zhuang et al.

Improving the performance of pre-trained policies through online reinforcement learning (RL) is a critical yet challenging topic. Existing online RL fine-tuning methods require continued training with offline pretrained Q-functions for stability and performance. However, these offline pretrained Q-functions commonly underestimate state-action pairs beyond the offline dataset due to the conservatism in most offline RL methods, which hinders further exploration when transitioning from the offline to the online setting. Additionally, this requirement limits their applicability in scenarios where only pre-trained policies are available but pre-trained Q-functions are absent, such as in imitation learning (IL) pre-training. To address these challenges, we propose a method for efficient online RL fine-tuning using solely the offline pre-trained policy, eliminating reliance on pre-trained Q-functions. We introduce PORL (Policy-Only Reinforcement Learning Fine-Tuning), which rapidly initializes the Q-function from scratch during the online phase to avoid detrimental pessimism. Our method not only achieves competitive performance with advanced offline-to-online RL algorithms and online RL approaches that leverage data or policies prior, but also pioneers a new path for directly fine-tuning behavior cloning (BC) policies.

Fengyuan Dai, Zifeng Zhuang, Yufei Huang et al.

Diffusion models have emerged as powerful generative tools across various domains, yet tailoring pre-trained models to exhibit specific desirable properties remains challenging. While reinforcement learning (RL) offers a promising solution,current methods struggle to simultaneously achieve stable, efficient fine-tuning and support non-differentiable rewards. Furthermore, their reliance on sparse rewards provides inadequate supervision during intermediate steps, often resulting in suboptimal generation quality. To address these limitations, dense and differentiable signals are required throughout the diffusion process. Hence, we propose VAlue-based Reinforced Diffusion (VARD): a novel approach that first learns a value function predicting expection of rewards from intermediate states, and subsequently uses this value function with KL regularization to provide dense supervision throughout the generation process. Our method maintains proximity to the pretrained model while enabling effective and stable training via backpropagation. Experimental results demonstrate that our approach facilitates better trajectory guidance, improves training efficiency and extends the applicability of RL to diffusion models optimized for complex, non-differentiable reward functions.

Zifeng Zhuang, Diyuan Shi, Runze Suo et al.

Complex high-dimensional spaces with high Degree-of-Freedom and complicated action spaces, such as humanoid robots equipped with dexterous hands, pose significant challenges for reinforcement learning (RL) algorithms, which need to wisely balance exploration and exploitation under limited sample budgets. In general, feasible regions for accomplishing tasks within complex high-dimensional spaces are exceedingly narrow. For instance, in the context of humanoid robot motion control, the vast majority of space corresponds to falling, while only a minuscule fraction corresponds to standing upright, which is conducive to the completion of downstream tasks. Once the robot explores into a potentially task-relevant region, it should place greater emphasis on the data within that region. Building on this insight, we propose the $\textbf{S}$elf-$\textbf{I}$mitative $\textbf{R}$einforcement $\textbf{L}$earning ($\textbf{SIRL}$) framework, where the RL algorithm also imitates potentially task-relevant trajectories. Specifically, trajectory return is utilized to determine its relevance to the task and an additional behavior cloning is adopted whose weight is dynamically adjusted based on the trajectory return. As a result, our proposed algorithm achieves 120% performance improvement on the challenging HumanoidBench with 5% extra computation overhead. With further visualization, we find the significant performance gain does lead to meaningful behavior improvement that several tasks are solved successfully.

Xing Lei, Zifeng Zhuang, Shentao Yang et al.

Recently, supervised learning (SL) methodology has emerged as an effective approach for offline reinforcement learning (RL) due to their simplicity, stability, and efficiency. However, recent studies show that SL methods lack the trajectory stitching capability, typically associated with temporal difference (TD)-based approaches. A question naturally surfaces: \textit{How can we endow SL methods with stitching capability and close its performance gap with TD learning?} To answer this question, we introduce $Q$-conditioned maximization supervised learning for offline goal-conditioned RL, which enhances SL with the stitching capability through $Q$-conditioned policy and $Q$-conditioned maximization. Concretely, we propose \textbf{G}oal-\textbf{C}onditioned \textbf{\textit{Rein}}forced \textbf{S}upervised \textbf{L}earning (\textbf{GC\textit{Rein}SL}), which consists of (1) estimating the $Q$-function by Normalizing Flows from the offline dataset and (2) finding the maximum $Q$-value within the data support by integrating $Q$-function maximization with Expectile Regression. In inference time, our policy chooses optimal actions based on such a maximum $Q$-value. Experimental results from stitching evaluations on offline RL datasets demonstrate that our method outperforms prior SL approaches with stitching capabilities and goal data augmentation techniques.