Weitao Zhou

Zhiming Xu, Weitao Zhou, Xianghui Pan et al.

Real-world dynamics shifts pose a critical challenge for reinforcement learning in robotics, as policies tightly coupled to nominal environments often fail catastrophically when physical conditions change. Most existing methods rely on encoding explicitly identified physical parameters into a latent context, a parameter-centric paradigm that depends on pre-specified axes of variation and becomes brittle under unmodeled or compound dynamics changes. We revisit dynamics adaptation from an outcome-centric perspective: rather than telling policies what the dynamics are, we enable them to learn how dynamics affect interaction outcomes. Theoretically, this is grounded in a monotonic relationship between target-domain regret and the Lipschitz constant of a trajectory dynamics encoder. Practically, this constant can be upper-bounded through contrastive learning, yielding a smooth, task-relevant latent topology without privileged dynamics information. On MuJoCo benchmarks, our method consistently outperforms parameter-centric baselines under severe dynamics shifts, including unmodeled and time-varying parameters, while also improving in-distribution stability and latent interpretability. Overall, these results validate that controlling latent geometry is a principled mechanism for robust adaptation.

Weitao Zhou, Zhong Cao, Yunkang Xu et al.

A typical trajectory planner of autonomous driving commonly relies on predicting the future behavior of surrounding obstacles. Recently, deep learning technology has been widely adopted to design prediction models due to their impressive performance. However, such models may fail in the "long-tail" driving cases where the training data is sparse or unavailable, leading to planner failures. To this end, this work proposes a trajectory planner to consider the prediction model uncertainty arising from insufficient data for safer performance. Firstly, an ensemble network structure estimates the prediction model's uncertainty due to insufficient training data. Then a trajectory planner is designed to consider the worst-case arising from prediction uncertainty. The results show that the proposed method can improve the safety of trajectory planning under the prediction uncertainty caused by insufficient data. At the same time, with sufficient data, the framework will not lead to overly conservative results. This technology helps to improve the safety and reliability of autonomous vehicles under the long-tail data distribution of the real world.

Zijian Zhang, Yuqing Jiang, Qian Cheng et al.

Vision-language-action (VLA) policies have advanced language-conditioned robotic manipulation by transferring semantic priors from pretrained vision-language models to action generation. Yet, standard action-imitation training often provides limited explicit supervision for 3D geometry, dense visual structure, and short-horizon environment evolution, which are critical for physically precise manipulation. We introduce \textbf{GaussianDream}, a feed-forward 3D Gaussian world-model plug-in that turns robot trajectories into structured spatial-temporal supervision. The key idea is to couple current Gaussian reconstruction with horizon-conditioned future Gaussian prediction during training, forcing a compact spatio-temporal prefix to be decodable into renderable 3D Gaussian states. This enables dense RGB rendering, depth, and pseudo 3D scene-flow supervision without requiring test-time Gaussian decoding. At inference, GaussianDream discards all auxiliary decoding heads and retains only the learned prefix to condition action generation, avoiding rendering, video rollout, or additional planning during closed-loop control. Experiments on LIBERO, RoboCasa Human-50, and real-robot tasks demonstrate strong and highly competitive performance, achieving \textbf{98.4\%} average success on LIBERO, \textbf{52.6\%} on RoboCasa Human-50, and \textbf{50.0\%} in real-world evaluation.

Bowen Jing, Ruiyang Hao, Weitao Zhou et al.

Generating safety-critical driving scenarios requires understanding why dangerous interactions arise, rather than merely forcing collisions. However, existing methods rely on heuristic adversarial agent selection and unstructured perturbations, lacking explicit modeling of interaction dependencies and thus exhibiting a realism--adversarial trade-off. We present CounterScene, a framework that endows closed-loop generative BEV world models with structured counterfactual reasoning for safety-critical scenario generation. Given a safe scene, CounterScene asks: what if the causally critical agent had behaved differently? To answer this, we introduce causal adversarial agent identification to identify the critical agent and classify conflict types, and develop a conflict-aware interactive world model in which a causal interaction graph is used to explicitly model dynamic inter-agent dependencies. Building on this structure, stage-adaptive counterfactual guidance performs minimal interventions on the identified agent, removing its spatial and temporal safety margins while allowing risk to emerge through natural interaction propagation. Extensive experiments on nuScenes demonstrate that CounterScene achieves the strongest adversarial effectiveness while maintaining superior trajectory realism across all horizons, improving long-horizon collision rate from 12.3% to 22.7% over the strongest baseline with better realism (ADE 1.88 vs.2.09). Notably, this advantage further widens over longer rollouts, and CounterScene generalizes zero-shot to nuPlan with state-of-the-art realism.

Qian Cheng, Weitao Zhou, Cheng Jing et al.

Real-world autonomous driving must adhere to complex human social rules that extend beyond legally codified traffic regulations. Many of these semantic constraints, such as yielding to emergency vehicles, complying with traffic officers' gestures, or stopping for school buses, are intuitive for humans yet difficult to encode explicitly. Although large vision-language models (VLMs) can interpret such semantics, their inference cost makes them impractical for real-time deployment. This work proposes LSRE, a Latent Semantic Rule Encoding framework that converts sparsely sampled VLM judgments into decision boundaries within the latent space of a recurrent world model. By encoding language-defined safety semantics into a lightweight latent classifier, LSRE enables real-time semantic risk assessment at 10 Hz without per-frame VLM queries. Experiments on six semantic-failure scenarios in CARLA demonstrate that LSRE attains semantic risk detection accuracy comparable to a large VLM baseline, while providing substantially earlier hazard anticipation and maintaining low computational latency. LSRE further generalizes to rarely seen semantic-similar test cases, indicating that language-guided latent classification offers an effective and deployable mechanism for semantic safety monitoring in autonomous driving.

Xiang Li, Gang Liu, Weitao Zhou et al.

Imitation learning (IL) has emerged as a central paradigm in autonomous driving. While IL excels in matching expert behavior in open-loop settings by minimizing per-step prediction errors, its performance degrades unexpectedly in closed-loop due to the gradual accumulation of small, often imperceptible errors over time.Over successive planning cycles, these errors compound, potentially resulting in severe failures.Current research efforts predominantly rely on increasingly sophisticated network architectures or high-fidelity training datasets to enhance the robustness of IL planners against error accumulation, focusing on the state-level robustness at a single time point. However, autonomous driving is inherently a continuous-time process, and leveraging the temporal scale to enhance robustness may provide a new perspective for addressing this issue.To this end, we propose a method termed Sequence of Experts (SoE), a temporal alternation policy that enhances closed-loop performance without increasing model size or data requirements. Our experiments on large-scale autonomous driving benchmarks nuPlan demonstrate that SoE method consistently and significantly improves the performance of all the evaluated models, and achieves state-of-the-art performance.This module may provide a key and widely applicable support for improving the training efficiency of autonomous driving models.

Weitao Zhou, Bo Zhang, Zhong Cao et al.

With the increasing presence of automated vehicles on open roads under driver supervision, disengagement cases are becoming more prevalent. While some data-driven planning systems attempt to directly utilize these disengagement cases for policy improvement, the inherent scarcity of disengagement data (often occurring as a single instances) restricts training effectiveness. Furthermore, some disengagement data should be excluded since the disengagement may not always come from the failure of driving policies, e.g. the driver may casually intervene for a while. To this end, this work proposes disengagement-reason-augmented reinforcement learning (DRARL), which enhances driving policy improvement process according to the reason of disengagement cases. Specifically, the reason of disengagement is identified by a out-of-distribution (OOD) state estimation model. When the reason doesn't exist, the case will be identified as a casual disengagement case, which doesn't require additional policy adjustment. Otherwise, the policy can be updated under a reason-augmented imagination environment, improving the policy performance of disengagement cases with similar reasons. The method is evaluated using real-world disengagement cases collected by autonomous driving robotaxi. Experimental results demonstrate that the method accurately identifies policy-related disengagement reasons, allowing the agent to handle both original and semantically similar cases through reason-augmented training. Furthermore, the approach prevents the agent from becoming overly conservative after policy adjustments. Overall, this work provides an efficient way to improve driving policy performance with disengagement cases.

Nanshan Deng, Weitao Zhou, Bo Zhang et al.

Current autonomous vehicles operate primarily within limited regions, but there is increasing demand for broader applications. However, as models scale, their limited capacity becomes a significant challenge for adapting to novel scenarios. It is increasingly difficult to improve models for new situations using a single monolithic model. To address this issue, we introduce the concept of dynamically enhancing a basic driving planner with local driving data, without permanently modifying the planner itself. This approach, termed the Dynamically Local-Enhancement (DLE) Planner, aims to improve the scalability of autonomous driving systems without significantly expanding the planner's size. Our approach introduces a position-varying Markov Decision Process formulation coupled with a graph neural network that extracts region-specific driving features from local observation data. The learned features describe the local behavior of the surrounding objects, which is then leveraged to enhance a basic reinforcement learning-based policy. We evaluated our approach in multiple scenarios and compared it with a one-for-all driving model. The results show that our method outperforms the baseline policy in both safety (collision rate) and average reward, while maintaining a lighter scale. This approach has the potential to benefit large-scale autonomous vehicles without the need for largely expanding on-device driving models.

Weitao Zhou, Zhong Cao, Nanshan Deng et al.

Self-driving vehicles (SDVs) are becoming reality but still suffer from "long-tail" challenges during natural driving: the SDVs will continually encounter rare, safety-critical cases that may not be included in the dataset they were trained. Some safety-assurance planners solve this problem by being conservative in all possible cases, which may significantly affect driving mobility. To this end, this work proposes a method to automatically adjust the conservative level according to each case's "long-tail" rate, named dynamically conservative planner (DCP). We first define the "long-tail" rate as an SDV's confidence to pass a driving case. The rate indicates the probability of safe-critical events and is estimated using the statistics bootstrapped method with historical data. Then, a reinforcement learning-based planner is designed to contain candidate policies with different conservative levels. The final policy is optimized based on the estimated "long-tail" rate. In this way, the DCP is designed to automatically adjust to be more conservative in low-confidence "long-tail" cases while keeping efficient otherwise. The DCP is evaluated in the CARLA simulator using driving cases with "long-tail" distributed training data. The results show that the DCP can accurately estimate the "long-tail" rate to identify potential risks. Based on the rate, the DCP automatically avoids potential collisions in "long-tail" cases using conservative decisions while not affecting the average velocity in other typical cases. Thus, the DCP is safer and more efficient than the baselines with fixed conservative levels, e.g., an always conservative planner. This work provides a technique to guarantee SDV's performance in unexpected driving cases without resorting to a global conservative setting, which contributes to solving the "long-tail" problem practically.

Weitao Zhou, Zhong Cao, Nanshan Deng et al.

Deep reinforcement learning (DRL) has emerged as a promising approach for developing more intelligent autonomous vehicles (AVs). A typical DRL application on AVs is to train a neural network-based driving policy. However, the black-box nature of neural networks can result in unpredictable decision failures, making such AVs unreliable. To this end, this work proposes a method to identify and protect unreliable decisions of a DRL driving policy. The basic idea is to estimate and constrain the policy's performance uncertainty, which quantifies potential performance drop due to insufficient training data or network fitting errors. By constraining the uncertainty, the DRL model's performance is always greater than that of a baseline policy. The uncertainty caused by insufficient data is estimated by the bootstrapped method. Then, the uncertainty caused by the network fitting error is estimated using an ensemble network. Finally, a baseline policy is added as the performance lower bound to avoid potential decision failures. The overall framework is called uncertainty-bound reinforcement learning (UBRL). The proposed UBRL is evaluated on DRL policies with different amounts of training data, taking an unprotected left-turn driving case as an example. The result shows that the UBRL method can identify potentially unreliable decisions of DRL policy. The UBRL guarantees to outperform baseline policy even when the DRL policy is not well-trained and has high uncertainty. Meanwhile, the performance of UBRL improves with more training data. Such a method is valuable for the DRL application on real-road driving and provides a metric to evaluate a DRL policy.