autonomous vehicle developers

Abstract: Significant progress has been made in safeguarding deep reinforcement learning (DRL) policies against input perturbations. Developing robust DRL involves three main stages: algorithm design, implementation, and evaluation. In this work, we identify and address a key limitation at each stage. First, we introduce Adversarial Importance Sampling (Advis), a method that uses importance sampling over trajectories from standard training to estimate and optimize verifiable worst-case returns. Advis satisfies three desirable criteria not jointly achieved by prior work: it requires no additional environment interactions, no auxiliary networks, and captures long-term robustness. Second, we introduce advrl, a modular PyTorch library that provides clean, single-file implementations of existing robustness methods and adversarial attacks, facilitating rapid prototyping and enabling reproducible and traceable evaluations. Third, we revisit evaluation under learned adversaries and show that optimal adversarial hyperparameters do not transfer across agents, which can lead to an overestimation of robustness when using a limited set of attacker configurations. Accordingly, we evaluate policies against a large and diverse set of attackers, using 6-14x more configurations than prior work. Finally, we evaluate our approach on continuous control environments, demonstrating its effectiveness relative to existing baselines. The code is available at: https://github.com/AmineAndam04/advrl

Abstract: We model total latency as the convolution of preprocessing, inference, and postprocessing distributions under a simplifying independence approximation, with selected stage parameters expressed as functions of source-image resolution. Under this assumption, the probability density of the total latency is the convolution of the stage-wise densities, and its cumulative distribution function (CDF) provides the distribution of end-to-end detection time. Each stage is modeled by a parametric distribution (e.g., Exponential, Erlang, Normal, Gamma), with parameters expressed as functions of the source-image resolution. Experiments with YOLOv11n on NVIDIA Jetson Orin NX using COCO2017 images across multiple resolutions assess the proposed models against fixed-parameter baselines using Kolmogorov Smirnov, Anderson Darling, and Cramér von Mises statistics. The results indicate that resolution-aware parameterization can improve distributional approximation in the measured setting, particularly for the more flexible Normal and Gamma models, while the quality of fit remains distribution dependent. Our contribution is a theoretically grounded and lightweight formulation for studying resolution-dependent latency distributions in a measured object detection pipeline.

Abstract: Datasets are a crucial element in the development of perception algorithms. They relate sensor measurement data to annotated reference information and allow for the deduction of sensor and object characteristics. In autonomous driving, the reference data commonly consist of semantic image segmentation, point-wise associations, or bounding box annotations. The dataset proposed in this work, however, aims to dig deeper into the evaluation of measurement principles and provides scanned 3D models of all vehicles together with a pose and continuous kinematics reference obtained by RTK-GNSS. Combined, the state of the complete dynamic surrounding of the sensor vehicle is known for any point in time. Subsequent reference formats can be easily computed in user-defined granularity. This dataset involves single-object and multi-object recordings with seven target vehicles. In particular, measurement effects such as occlusion, as well as reflections, can be evaluated, as the normals of the shape of the target vehicles are known. We describe the dataset, discuss the technical background of its development, and briefly present exemplary evaluations.

Abstract: Infrared-visible image fusion facilitates robust multimodal perception by integrating complementary textural nuances from visible sensors with thermal signatures from infrared systems. Due to the task's inherently ill-posed nature, existing methods heavily rely on structural priors but typically enforce rotation equivariance uniformly across all features. Such a holistic approach overlooks a critical distinction where low-frequency shared structures strictly adhere to equivariant constraints while high-frequency modality-specific details require greater flexibility to preserve unique information. To bridge this gap, we propose RoES, a Rotational Equivariant Selective-frequency fusion network. Instead of employing static decomposition, we introduce a trainable rotation-enhanced updater/predictor module to dynamically decouple low- and high-frequency components. The resulting representations are then processed through a dual-branch fusion module tailored for spectral consistency. Specifically, a rotation-equivariant Mamba is employed to capture long-range structural dependencies in the low-frequency domain, while a polar spectral attention-based Dual-Fourier block refines high-frequency details under explicit low-frequency guidance. Extensive experiments demonstrate that RoES consistently achieves state-of-the-art performance in both fusion quality and downstream object detection, establishing a robust solution for multimodal fusion by reconciling frequency-selective features with equivariant constraints. The source code is available at https://github.com/BryceLosky/RoES-Fusion.

Abstract: We consider an agent acting against a general non-Markovian environment. The agent maintains its agent states, but is free to choose a transition kernel across those states and optimize its state-feedback control policies. We study the bi-level agent design problem that optimizes the transition kernel and the policy it induces, given said kernel with offline data of observations and actions obtained via a behavioral policy. For general environments, we show that a soft $Q$-learning algorithm converges almost surely to the fixed point of a soft Bellman equation defined by the stationary averages that the behavioral policy and the chosen kernel induce, and we delineate what separates the resulting policy from an optimal one. In partially observed Markov decision problems, we analyze convergence properties of parametrized transition kernel design via zero-th order and Bayesian optimization techniques.

Abstract: Multiview point cloud registration is particularly challenging in low-overlap scenes, where reliable correspondences are limited and incorrect pairwise transformations can affect global pose estimation. In addition, registering all scan pairs is computationally expensive because many pairs provide weak geometric information. To address these problems, we propose GMPCR, a non-learning-based spectral consistency-guided framework for efficient and robust multiview point cloud registration. GMPCR builds a refined second-order compatibility structure from initial correspondences and uses its dominant spectral response to evaluate both correspondence reliability and scan-pair confidence. This allows unreliable correspondences to be filtered and informative scan pairs to be selected before relative transformation estimation, leading to a sparse pose graph and reduced pairwise registration cost. For each retained scan pair, maximal-clique-based hypothesis generation is used to estimate reliable relative transformations. The resulting pose graph is further refined by an adaptive history-aware synchronization scheme, in which the effect of residual history is adjusted according to changes in the global rotation residual. A recovery mechanism also allows down-weighted edges to regain confidence when their global consistency improves. Experiments on 3DMatch, 3DLoMatch, ScanNet, and ETH demonstrate the effectiveness of GMPCR. It achieves registration recalls of 97.2% and 89.6% on 3DMatch and 3DLoMatch, respectively, while maintaining competitive performance on ScanNet and ETH. The results show that GMPCR provides a favorable balance among registration accuracy, robustness to low overlap, and computational efficiency. The code is publicly available at https://github.com/swccj/gmpcr.

Abstract: Humanoid robots must traverse cluttered obstacle fields using onboard proprioceptive and visual observations, yet existing methods usually process multimodal observations without explicitly considering their different characteristics: proprioceptive observations are low-dimensional but governed by highly nonlinear robot dynamics, while egocentric visual observations are high-dimensional, noisy, and redundant. We propose DWMP (Dual World Model Policy), a framework that provides the actor with separate but complementary world-model representations for humanoid obstacle traversal. A Koopman-based dynamics world model lifts proprioceptive observations into a latent space where their temporal evolution is approximately linear, making the dynamics features easier for the actor to learn from. An RSSM-based visual world model compresses egocentric depth observations into compact stochastic states while preserving obstacle-related geometry. The student policy receives the fused latent representation for action generation, combining linearized proprioceptive dynamics with compressed visual perception. Experiments in simulation and on a Unitree G1 humanoid robot show that DWMP improves obstacle traversal performance over baselines and supports real-world deployment under randomized obstacle layouts.

Abstract: Multi-object tracking (MOT) is dominated by the tracking-by-detection paradigm, whose methods typically rely on a small set of hyperparameters that are conventionally chosen by hand. Tuning them requires repeated expert-guided experimentation, while the procedures used to select reported values are often not systematically evaluated or fully documented. Hyperparameter optimization (HPO) automates this process, yet it remains rarely used in MOT, and existing studies applying HPO to MOT predate modern deep-detector-based trackers and HOTA evaluation. We systematically apply HPO across two datasets and four tracking-by-detection methods. We also propose Multi-Fidelity Greedy Coordinate Search (MFGCS), which optimizes one hyperparameter at a time by first evaluating candidate values on a small subset of scenes and re-evaluating only promising candidates on the full dataset. Across all eight tracker-dataset combinations, the Tree-structured Parzen Estimator (TPE) and MFGCS outperform both our hand-tuned configurations and the corresponding published results, with improvements of up to 4.38 and 16.05 HOTA points, respectively. MFGCS also reaches a predefined HOTA target faster than TPE in seven of the eight combinations. Within each tracker-dataset pair, all optimizers share the same search space and evaluation pipeline, isolating the effect of the search strategy. We release the code and tuned configurations to enable future work to compare against systematically optimized rather than default or manually tuned baselines.

Abstract: The development of autonomous driving demands comprehensive testing in mixed-traffic scenarios involving vulnerable road users (VRUs), where purely artificial agents often fail to capture authentic human social negotiations. While human-in-the-loop (HITL) simulators enable safe investigation of these interactions, existing multi-agent platforms struggle with the network latency and synchronization constraints required for high-fidelity haptic feedback. To resolve this, we present CARLAverse, an open-source, multimodal simulation ecosystem. Extending modular hardware abstraction, CARLAverse integrates driving (DrivoCARLA), cycling (CycloCARLA), and pedestrian (WalkoCARLA) simulators into a shared virtual environment. Its core methodological contribution is a distributed physics architecture: latency-critical ego dynamics and high-frequency force feedback are computed locally on client nodes, while a central CARLA server orchestrates non-player character (NPC) physics and global traffic. By decoupling haptic control loops from network bottlenecks, CARLAverse enables scalable, cross-institutional HITL experiments without compromising physical immersion. Code and documentation: https://git.ieem-ka.de/simulator-environments/carlaverse

Abstract: With the rising demand to decouple objects from 3D scenes, we propose SAMV-DUSt3R, an end-to-end model that injects SAM2 2D masks into MV-DUSt3R reconstruction. A Cross Flow Mask Block uses these masks to steer the network toward the target instance, jointly improving shape accuracy and achieving object-level disentanglement without multi-stage pipelines. To ensure reconstruction stability, a lightweight Spatial RankGNN selects the optimal reference view with a selection accuracy of 73.5\%. Extensive experiments demonstrate that our method boosts average reconstruction precision by 11\% across various metrics compared to state-of-the-art baselines. These results reveal a strong instance-disentanglement capability and clear benefits for driving, robotics, AR/VR, and heritage digitisation.

Abstract: Science-inspired vision models show that explicit propagation dynamics can provide structured and interpretable alternatives to conventional token mixing. Existing formulations, however, typically construct and adapt visual propagation within a particular dynamical family, while visual representations can require substantially different spatial interactions across samples, channels, and network stages. We explore cross-regime adaptive propagation and introduce TailProp, a hierarchical vision backbone built upon the Tail Propagation Operator (TPO). TPO uses Gaussian and Cauchy stable-process propagators as complementary bases with rapidly decaying and heavy-tailed spatial influence, and predicts a content-conditioned channel-wise coefficient to adaptively combine them. Because this coefficient is spatially shared, the two responses are fused directly in the DCT domain with a single DCT/IDCT pair, yielding $O(N^{1.5})$ spatial mixing for square feature maps with $N=HW$ and fixed channel width. Across image classification, object detection, semantic segmentation, robustness, and cross-backbone restoration, TailProp consistently outperforms matched propagation baselines; TailProp-B reaches 84.4% Top-1 accuracy on ImageNet-1K, 50.3/44.8 box/mask AP under the 3x Mask R-CNN schedule, and 50.8% mIoU on ADE20K. Controlled ablations further show that these gains are not explained by single-basis propagation, an additional same-family branch, or within-family adaptive order alone, supporting complementary two-basis propagation as an effective design principle for visual representation learning.

Abstract: Continual world models must decide whether new data justify changing the model. Fixed replay schedules and prediction-error triggers specify when to update, but neither reveals the value of an individual update: one deployment run cannot show how the same model would have performed at that moment had it held its parameters. We introduce the fork ledger, which branches a deployment stream at pre-registered decision points into matched update and hold continuations under common random numbers. It evaluates both continuations on the same episodes and records $ΔR = R_{\mathrm{update}} - R_{\mathrm{hold}}$. Always applying one fixed update mechanism lowers return on all three simulated control tasks: CartPole ($-144.0$; checkpoint-bootstrap $95\%$ CI $[-185.4,-116.1]$, against a converged return near $650$), Walker ($-82.8$; $[-101.1,-61.7]$) and Cheetah ($-18.6$; $[-29.0,-6.6]$). Divergence is an outcome of applying the update, so the estimand counts every attempted fork; restricted to the $693$ of $720$ that did not collapse, CartPole and Walker are unchanged in sign ($-113.4$ and $-82.1$) and Cheetah becomes unresolved ($-3.9$; $[-17.5,+13.0]$). The task is the unit of inference: each contributes $240$ attempted forks over five pretrained checkpoints crossed with two drift directions. The ledger makes counterfactual utility observable for a fixed mechanism, allowing triggers to be judged by the updates they select rather than by surprise detection alone.

Abstract: Reinforcement learning (RL) agents deployed in real-world environments are often vulnerable to adversarial perturbations in state observations, creating risks in safety-critical applications. Certification methods can improve robustness against adversarial perturbations by providing lower bounds on expected cumulative rewards. Existing certification methods, however, mainly focus on risk-neutral objectives. In this paper, we extend certification methods to risk-sensitive objectives by establishing lower bounds on the exponential utility of cumulative rewards under $l_{p}$-norm-bounded state adversarial perturbations ($1\leq p <\infty$). By introducing a $φ$-divergence relaxation of the perturbation set, we formulate the risk-sensitive certification problem as a convex optimization and derive its dual to obtain a tractable approximation of the certified lower bound. We further propose an empirical method that improves certified lower bounds by selecting the training risk-aversion parameter $β$ independently of the risk level used during evaluation. Experiments on both OpenAI Gym environments and a machine replacement problem show that, compared to risk-neutral training, risk-averse training generally yields policies with higher certified lower bounds, particularly under larger perturbation budgets. Moreover, under both risk-neutral and risk-averse evaluation settings, increasing risk aversion during training leads to non-monotonic certification performance, where certified lower bounds initially improve but eventually decrease due to overly conservative policies.

Abstract: The pedestrian crossing intention task involves predicting whether pedestrians are likely to cross the road from the point of view of an autonomous vehicle. We introduce TrajFusionNet+, a novel transformer-based model for pedestrian crossing intention prediction. TrajFusionNet+ combines sequential and visual representations of pedestrian trajectory with a graph-based representation of the scene context in order to predict pedestrian crossing intention. The proposed architecture builds upon our previous model, TrajFusionNet, and comprises three branches: a Sequence Attention Module (SAM), which processes a sequential representation of past and predicted pedestrian trajectories; a Visual Attention Module (VAM), which utilizes a visual representation of the pedestrian trajectories by overlaying observed and predicted bounding boxes onto scene images; and a Graph Attention Module (GAM), which extracts pedestrian-centric graphs from segmented scene images and captures the relational dependencies between pedestrians and traffic elements. TrajFusionNet+ achieves improved state-of-the-art performance on the two most widely used pedestrian crossing intention datasets, PIE and JAAD. Furthermore, we introduce a new evaluation protocol in which models are trained jointly on the PIE and JAAD datasets but evaluated separately on each. Under this setting, TrajFusionNet+ demonstrates superior generalization compared to existing approaches.

Abstract: Accurate estimation of lateral vehicle dynamics near the adhesion limit is important for stability control and high-performance driving, but lateral velocity is rarely measured directly because sensors such as optical sensors are costly. This paper presents an inertial measurement unit (IMU)-centric Moving Horizon Estimation framework that reconstructs lateral velocity using standard onboard signals, without relying on exteroceptive odometry or detailed tire-parameter tuning. Experimental validation on human-driven sports cars and an autonomous open-wheel race car across tracks, maneuvers, and conditions demonstrates accurate and robust lateral velocity and lateral acceleration estimates. The proposed framework is available at https://github.com/Aseuffo/IMU-Centric-MHE

Abstract: Computing the closest point on a clothoid is a recurring task in geometric design, road and railway alignment, and path planning. The efficient algorithm of Frego and Bertolazzi addresses this problem, but its candidate-selection analysis assumes at most one local minimum per search interval. We exhibit admissible configurations with two local minima, raising the question of whether the existing strategy accounts for every possible minimum. Using the geometry of the clothoid evolute, we prove that, for any query point and any proper no-inflection planar clothoid segment with tangent-angle variation at most $2π$, the squared-distance function has at most three stationary points; if all three are local extrema, their order is min-max-min. This establishes the completeness of the original candidate-selection logic beyond the one-minimum premise. It also shows that no interior search is needed when neither endpoint derivative test is active, allowing unnecessary midpoint searches to be omitted while retaining numerical fallback. Numerical experiments demonstrate reductions in iteration count and evaluation time.

Abstract: UAV target following in cluttered environments is inherently predictive: current-state followers can lag behind turns, choose blocked corridors, or trade tracking for unsafe near-horizon motion. We propose a future-aware flow planning framework for state-informed UAV target following. Predicted target futures guide clean UAV trajectory generation as horizon-aligned residual signals, while risk-scored executable-prefix repair is embedded inside the sampling loop. On fixed ID/OOD receding-horizon benchmarks, the planner improves the intended safety--tracking trade-off rather than dominating every metric: it matches zero measured ID collision rate with the highest ID safe-tracking time, and gives the lowest OOD macro collision rate and final tracking error among the displayed methods, while Future-MPC remains smoother and stronger on some thresholded OOD success metrics under its hand-designed objective. Ablations show that future adaptation improves candidate generation before safety repair, and simulator-facing stress tests probe interface, sensing, and controller-execution effects. These results support horizon-aligned future adaptation and embedded prefix repair as complementary ingredients for safe UAV target following under the tested simulation conditions.

Abstract: Driver gaze provides information regarding driver visual attention and situational awareness to the surrounding traffic. Existing driver gaze estimation studies represent gaze in terms of gaze zone or gaze vector/point-of-gaze (PoG). However, object-level gaze information provides a more semantically meaningful representation of visual attention by identifying attended objects, such as vehicles, pedestrians, or traffic signals. In this study, we propose an end-to-end driver gaze object prediction framework, TransGaze-Object, Transformer-based Gaze Object prediction model. The proposed framework first extracts facial features, including face and iris-weighted eye features, along with trafficobject spatial features. A transformer based cross-attention mechanism is then used to compute similarity scores and attention weights for predicting the drivers gaze object. To train this model, we propose a benchmark driver gaze dataset, Urban Driving-Face Scene Gaze (UD-FSG), comprising synchronized driver-face and traffic-scene images, scene objects bounding boxes, and gaze labels in terms of 2D gaze coordinate and gaze object. The TransGaze-Object model achieves an overall accuracy of 60% for gaze-object prediction, compared to 51% accuracy obtained from associating the estimated Point-of-Gaze to traffic objects. The error analysis reveals that TransGaze-Object reduces confusion between traffic objects (predicted) and the background (ground-truth), achieving an error rate of 11.68%, a 49.7% relative reduction compared with 23.21% error obtained from PoG-based gaze-object association. Overall, the results demonstrate the effectiveness of directly predicting gaze objects from driver-face and traffic-scene information, rather than estimating an intermediate Point-of-Gaze and subsequently associating it with traffic objects.

Abstract: Modern vision-language-action (VLA) policies predict a whole chunk of actions: one to two seconds of coordinated motion emitted in a single forward pass. Yet an action chunk is essentially a short multivariate trajectory, but inside these models it is a sequence of generic per-timestep hidden tokens decoded by a linear head. This under-serves two motion structures. First, frequency: a chunk superimposes a smooth global trend and fine corrective motion across time scales, and a single token entangles them. Second, cross-phase geometry: motions of different phases (reach, contact, grasp adjustment, settling) unfold along very different, near-orthogonal directions in representation space, yet are tightly related for the task and arise across the time axis. Dot-product attention scores alignment by an inner product, so it favors aligned tokens and is least sensitive near orthogonality, leaving such relationships for the network to recover through a detour. We introduce Time-Frequency Geometric Cross-Attention (TFGCA), a drop-in module repairing both blind spots. TFGCA uses a per-dimension learnable stationary wavelet transform to decompose the action chunk into time-frequency tokens, and each time token retrieves information from them via a cross-attention that fuses the dot product (similarity) with the wedge-product magnitude (sensitive to near-orthogonality) through a learnable weight. A zero-initialized residual reproduces the base behavior at initialization, so it can be dropped onto a pretrained VLA and fine-tuned jointly. Relative to the same-source base, TFGCA improves in-distribution LIBERO by +1.5 on average, the OOD LIBERO-Plus by +6.3, the randomized average under RoboTwin domain randomization by +28.5, and the overall success rate on three real-robot AgiBot A2 tasks by +11.67 points, with larger gains out of distribution.

Abstract: Test-Time Adaptation (TTA) enhances the robustness of models against distribution shifts but faces two critical challenges: error accumulation from noisy pseudo-labels and catastrophic forgetting of source knowledge. Uncertainty-based approaches designed to mitigate error accumulation often yield overconfident or computationally expensive estimates, while strategies intended to prevent forgetting via prototype replay rely on static representations that easily become misaligned as the model adapts. To address these issues, this paper proposes a novel framework, Distilling Image Prototype for Guided Test-Time Adaptation (DIPTTA). The core of the proposed approach is the introduction of a Distill Image Prototype (DIP), a compact set of synthetic images that serves as a dynamic and regenerative anchor of source knowledge. This prototype enables a dynamic feature replay mechanism that continuously generates feature prototypes aligned with the current state of the model, thus effectively preventing catastrophic forgetting. Furthermore, the DIP anchors a source-calibrated uncertainty estimation method, which provides a less biased measure of sample reliability by leveraging stable source knowledge, thereby robustly suppressing error accumulation. Extensive experiments on multiple benchmarks demonstrate that DIPTTA significantly outperforms state-of-the-art methods, particularly under severe domain shifts. The source code is available at https://github.com/LiwenWang919/DIPTTA.

Abstract: Sampling-Based Model-Predictive Control (MPC) algorithms are a flexible class of controllers used for navigation on a wide range of robotic systems. Historically, such approaches have lacked hard safety guarantees, a shortcoming which we remedy in this work by computing guaranteed reachable-set overapproximations online with a fast, interval-based pipeline. We show that our method achieves similar performance to a state-of-the-art reachability-based planner without the need for the expensive pre-computation step, and can be scaled to systems that are infeasible using existing approaches. Finally, we demonstrate that our technique reduces safety violations by over 99% in a racing simulation and successfully controls a model racecar on real hardware experiments without crashes.

Abstract: Graph-based surrogate models offer a promising route to accelerate computational fluid dynamics (CFD) simulations on unstructured meshes. However, their development is limited by the scarcity of benchmark datasets spanning multiple flow regimes and standardized protocols for long-horizon autoregressive prediction. We introduce ONECYL (ONE CYLinder), a new benchmark for unsteady flow past a circular cylinder across laminar, transitional, and high-Reynolds-number regimes. The benchmark comprises 450 high-fidelity Variational Multiscale finite-element simulations (270,000 flow snapshots) with randomized cylinder geometries, providing time-resolved velocity and pressure fields together with mesh connectivity, geometric descriptors, Reynolds numbers, and integrated aerodynamic quantities. Beyond the dataset, ONECYL establishes a unified evaluation framework combining full-field rollout errors, virtual probes, and drag and lift predictions to assess numerical accuracy and physical fidelity. To accompany the benchmark, we develop a Graph Transformer as a reference baseline predicting velocity and pressure fields autoregressively on unstructured meshes. Using ONECYL, we investigate geometric representations and physics-based regularization across the three Reynolds-number regimes. The results show that explicitly encoding the cylinder geometry through a level-set representation consistently improves long-horizon prediction accuracy and generalization to unseen geometries, while divergence-based regularization becomes increasingly beneficial as flow complexity increases. The ONECYL benchmark and its Graph Transformer baseline provide a reproducible framework for evaluating graph-based surrogate models and establish a foundation for future research on long-horizon prediction of unsteady bluff-body flows.

Abstract: Multi-agent traffic simulation seeks diverse, coordinated, and physically realistic futures from maps and observed history. Long-horizon closed-loop generation must reconcile multiple decision time scales while its context evolves with generated states. Existing methods often unfold long futures from the initial scene and resolve intent, interaction, and motion monolithically, weakening cross-scale consistency and adaptation. We present HI-FLOOP, a branch-consistent multi-timescale state-feedback framework. Eight scene-level Worlds represent joint hypotheses, and all agents share the selected World identity throughout an 8-second rollout. Within the branch, an 8-second Goal anchors intent, a 2-second Preview coordinates interactions, and 1-second Control produces physical motion. Every 0.5-second commit feeds back only its executed prefix as new facts, while unexecuted hypotheses never enter factual memory. Joint Preview Interaction (JPI) induces a sparse directed future graph from Preview and uses conflict probabilities and signed arrival-time differences to gate interaction refinement. For generated-state recovery, a prefix-frozen A-to-B cascade lets frozen Model A generate 0-1 seconds, then transfers typed physical state, admissible context, and the branch index, but no latent state, to an independent Model B for re-encoding and 1-2-second recovery. On the full H-D public-validation split of 955 scenarios, one complete S1 run yields an Overall score of 0.689987 with the official evaluator. Under agent-centric oracle evaluation, HI-FLOOP achieves oracle-minADE@8 of 1.196636 m over the 8-second horizon and 0.526 m over the 6-second horizon.

Abstract: Autonomous navigation in unstructured environments requires robust scene understanding, yet Vision-Language Models (VLMs) often suffer from semantic ambiguity, where conflicting predictions can lead to dangerous failures. To address this, we present a novel pipeline for vision-based traversability estimation that explicitly models contextual uncertainty. Our approach utilizes Conceptual Anchoring to ground open-vocabulary VLM predictions onto a continuous physical traversability scale. By formulating the model's responses as a Gaussian Context Distribution (GCD), we derive both a dense traversability map and a dense uncertainty map based on the statistical properties of the distribution. Experimental validation on the real-world GOOSE dataset demonstrates that our proposed uncertainty metric effectively correlates with sources of ambiguity, such as visual artifacts and mixed terrain overlap. The method exhibits competitive performance while offering the distinct advantage of providing statistical uncertainty estimates to address semantic ambiguity, enabling safer and more reliable autonomous behavior in complex outdoor settings.

Abstract: Several geometry-aware approaches to low-rank adaptation have emerged for parameter-efficient fine-tuning of large pre-trained models. These methods aim to take full advantage of the geometric structure of low-rank manifolds for improving the efficiency in subspace utilization and reducing redundancy by enforcing orthogonality constraints during optimization. The strong empirical results of these techniques have motivated further study into whether predictions from such geometry-based adaptation methods could be overconfident. In this paper, we build on the singular value decomposition factorization of adapters to develop a framework based on Stein variational gradient descent (SVGD). In this formulation, the low-rank matrices are transported along the Stiefel manifold to match the targeted distributions while retaining their crucial geometric structure. Since this geometry-aware SVGD approach provides multiple solutions during inference, it supports uncertainty quantification and produces better-calibrated adapters on the Stiefel manifold. Extensive experiments show that our method delivers strong model calibration and attains higher prediction accuracy than SVGD and related uncertainty estimation methods that are formulated in Euclidean space.

Abstract: Object detection and tracking are fundamental components of perception systems for autonomous driving. Achieving robust performance under adverse conditions such as limited visibility, sensor noise, and failures remains an open challenge, particularly in autonomous racing, where vehicles operate at very high speeds, experience strong vibrations, and interact under small safety margins. This paper presents a multi-modal late-fusion perception pipeline for object detection and tracking in the autonomous racing domain. The proposed system extends previous work by exploiting all onboard sensors through a late-fusion approach and a dedicated multi-object tracking framework. Independent detections from cameras, LiDARs, and RADARs are combined to provide timely and robust state estimates of surrounding vehicles. The tracking method explicitly compensates for detection delays and embeds in its model prior knowledge of vehicle dynamics and track layout. Experimental evaluation on real-world data across diverse critical scenarios, representative of challenging edge cases also in urban driving, confirms the effectiveness of the proposed pipeline and its suitability to support safe and adaptive planning decisions.

Abstract: Vision-language models (VLMs) demonstrate strong performance across compositional reasoning benchmarks, which require reasoning over semantic perturbations of objects, attributes, relations, and their interactions. However, our controlled analysis reveals that existing compositionality-aware VLMs exhibit element-specific biases, often underperforming vanilla CLIP on certain compositional elements. To address this, we propose Compositional Scene Graph-guided CLIP (CS-CLIP), which uses scene graphs to identify compositional elements and construct structured negatives via selective masking. We further retain negatives that are most contradictory to the original caption, forcing the model to rely on compositional structure rather than surface cues. CS-CLIP achieves state-of-the-art compositional reasoning with robust performance across compositional elements. It also preserves general vision-language capabilities such as cross-modal retrieval and downstream visual reasoning, while requiring fewer training samples than prior methods.

Abstract: Ensuring safety in reinforcement learning under nonstationarity requires anticipating changes in risk before they lead to unsafe behavior. Existing approaches typically rely on safety constraints defined at design time or updated reactively during execution, assuming that such constraints remain valid over time. However, in nonstationary environments with evolving contexts and changing driving layouts, these assumptions may fail. We propose a framework for proactive safety constraint generation based on context forecasting. The approach infers latent environmental context from observations, predicts its future evolution, and constructs safety constraints adapted to anticipated conditions. This enables the agent to proactively avoid unsafe regions instead of reacting only after safety violations occur. We evaluate the method in driving environments with structured context variation. The experiments include a sweep over nonstationarity intensities and additional held-out driving layouts, including highway, intersection, and racetrack scenarios. Results show that proactive constraint generation substantially reduces collisions under both seen and out-of-training nonstationarity intensities and generally remains effective across held-out driving layouts while maintaining usable task performance. These findings suggest that context-based constraint generation is a promising approach for safe reinforcement learning under nonstationarity.

Abstract: We study the control of Markov decision processes in which the quality of a policy is evaluated by a dynamic, time-consistent Markov risk measure rather than by an expected discounted cost. The main obstacle to combining such measures with reinforcement learning is that a transition risk mapping depends on the transition kernel in a nonlinear way, and therefore cannot be estimated from a single observed transition. We remove this obstacle by employing mini-batch transition risk mappings: the mapping is applied to the empirical measure of $N$ independent next-state samples, and the result is averaged. The resulting mapping is again coherent. However, as an expected value of a function of $N$ next-state values, it admits an unbiased one-sample estimator. We embed this mapping into a double deep Q-network, analyze the two sources of estimation bias that arise, and obtain a risk-averse Q-learning method applicable to state spaces far beyond the reach of tabular schemes. The method is applied to an underwater robot navigation problem, in which a vehicle must visit collection points, gather stochastic information payloads, and deliver them at transmission points, while exposed at each step to the risk of destruction. A hierarchical decomposition delegates path execution to an exact graph search and confines learning to the high-level ``collect or transmit'' decision. A low-dimensional feature map, invariant under the symmetries of the problem, replaces the raw state--configuration encoding. In experiments on $300$ held-out environments, the resulting policies transfer to instance sizes never seen in training, and already $N=2$ reduces the upper semideviation of the outcome distribution while simultaneously improving its mean whenever the simulator is misspecified---an empirical counterpart of the duality between coherent risk measures and distributional robustness.

Abstract: Multimodal Large Language Models (MLLMs) show strong progress on vision-language tasks, yet their reliability in safety-critical settings remains underexplored. Fire-smoke understanding is central to public safety and disaster response, but most existing benchmarks lack diverse real-world scenarios and context-aware evaluation. We introduce SAFIRE, a large-scale benchmark for fire-smoke understanding in MLLMs, comprising 83K captioned images from 20 scenarios and 193K multiple-choice VQA (MCVQA) generated from a 9.7K-image subset, spanning 10 evaluation dimensions from basic perception to higher-order reasoning. A GPT-5.4-assisted multi-stage verification pipeline with MLLM majority voting ensures annotation quality. Evaluating ten open-source MLLMs (8B-38B) yields an average accuracy of 61.9%, exposing major gaps in safety-critical reasoning. We further show that adapting vision encoders with only 7% of our domain-specific data boosts fire-scene classification accuracy from 20.1% to 64.5%, indicating that carefully curated data can yield substantial gains even when data volume is limited. All datasets, models, and code are available at https://risys-lab.github.io/SAFIRE/.

Abstract: LiDAR-based 3D Single Object Tracking (3D SOT) is critical for robotic perception and navigation and aims to localize dynamic objects across frames in sparse point clouds. Existing methods, rooted in the Siamese tracking paradigm from 2D vision, rely on costly dual-input designs and excessive motion modeling guided by template priors, hindering their efficiency. Our in-depth analysis reveals: (i) the template paradigm is redundant, as the previous bounding box center encodes sufficient historical context; (ii) complex motion modeling is unnecessary, as geometric alignment provides adequate motion priors. Based on the above findings, we propose the first Template-Free Tracking framework (TFTrack). The novel framework eliminates the need for template-search pairings and operates directly on the current frame guided solely by the prior bounding box center and size. We instantiate this paradigm into three variants: TFTrack-Voxel, TFTrack-Pillar, and TFTrack-Point, to explore different 3D representations under a unified framework, ensuring flexibility across sparse and dense scenes. Extensive experiments on KITTI and nuScenes benchmarks show that TFTrack is competitive with leading template-based trackers, while reducing FLOPs by approximately 50% and running at approximately 120 FPS. By simplifying overcomplicated motion-centric designs, TFTrack establishes a new minimalist paradigm for efficient 3D point cloud tracking, paving the way for real-time and resource-efficient deployment in embedded robotic systems, such as autonomous vehicles. The code is available at https://github.com/tftrack-anonymous/TFTrack/tree/main.

Abstract: Bridging the simulation-to-reality gap in roadside LiDAR requires addressing several coupled discrepancies, including scene geometry, sampling density, return patterns, and pedestrian scale. This report presents a multi-source collaborative training and class-aware fusion framework for Sim2Real 3D detection. The method organizes digital-twin scans, diffusion-redrawn scans, density-stabilized scans, and pedestrian morphology-aligned samples into a unified training pool with complementary roles. Within a common DSVT detection formulation, source-specialized expert branches preserve those roles while optimizing for the same detection objective. At inference, a predefined class-aware fusion pathway integrates geometry-stable and calibration-aware branches for vehicles, sampling-complementary branches for trucks, and morphology-consistent evidence for pedestrians. A label-free point-cloud center blend then refines geometric localization. On the UrbanTwin V2X-Real hidden test set, the unified system achieves a combined score of 0.7421, with 3D mAP@0.5 of 0.4518 and a realism score of 0.8871. The results indicate that a stable, interpretable collaboration among data sources is more valuable than unconstrained aggregation of model outputs.

Abstract: Target-based LiDAR-camera extrinsic calibration is a prerequisite for multi-sensor fusion in robotics. However, in the widely adopted four-hole pipeline, calibration accuracy is bottlenecked by LiDAR-side hole-center extraction, which suffers from sparse angular coverage and mixed-pixel corruption. This paper presents P$^2$Calib, which exploits pattern priors, geometric constraints specified by the CAD model of the target board, to improve calibration accuracy. First, we incorporate the known hole radius as a fitting constraint to prevent center estimates from degrading under sparse angular coverage. Building on the improved hole estimates, we further enforce the rigid rectangular layout of the four holes as a global consistency constraint to correct residual errors across holes. Both priors are integrated into an interactive calibration tool that provides a complete extrinsic calibration pipeline. Experiments on simulated and real datasets show that P$^2$Calib lowers the joint registration residual by 90\% and 82\% and the held-out reprojection error by 96\% and 77\% over the baseline. Code, https://github.com/JokerJohn/P2Calib.git, and data will be released to facilitate future research.

Abstract: Standard Convolutional Neural Networks (CNNs) exhibit severe performance degradation due to a strong inductive texture bias that prioritizes local, high-frequency patterns over global structural shapes. This dependency causes confident misclassifications during textural changes or environmental effects. To address this flaw, this study introduces the Texture-Penalized Prototype Network (TPPN), a novel architectural framework that shifts this inherent bias without depending on resource-intensive augmented datasets. Specifically, a Texture-Penalization Branch (TPB) imposes a penalty to suppress the extraction of local texture proxies, forcing the network backbone to discard high-frequency cues and extract purified, shape-biased representations. By evaluating similarities within a prototype-based hypersphere derived from the final convolutional features, the approach enforces strict geometric constraints, treating objects as compositions of essential parts to achieve robust classification. Evaluations on texture-shape cue-conflict datasets and synthetic noise benchmarks demonstrate the stronger shape bias of this structural disentanglement. The proposed framework reduces the inherent texture bias of a baseline ResNet-50 from 55.11% to 29.73%, surpassing the texture-suppression capabilities of an off-the-shelf Vision Transformer (ViT-B/16). Furthermore, the approach demonstrates robust generalization under cue-conflict conditions, resisting textural shortcut learning when encountering Out-of-Distribution (OOD) shapes. The model maintains stronger shape accuracy against elevated perturbations. On clean validation data, the architecture incurs a minimal drop in accuracy of 0.90 percentage points. This provides a structural, efficient solution to CNN texture bias.

Abstract: In decentralized decision-making, several agents connected according to a network graph aim at solving a classification problem by collecting streaming observations. Due to decentralization, they run an iterative algorithm where, at each iteration, they can only exchange information locally with their neighbors. While decentralized estimation solutions have been shown to match the performance of optimal centralized systems, we show here that surprisingly this conclusion does not hold for decentralized decision-making. Specifically, we prove that the error probability for the best decentralized decision strategy exhibits an irreducible loss with respect to the optimal centralized classifier. This result establishes a fundamental limit for the performance of any decentralized decision strategy. We obtain an analytical relation showing that this limit is related to the interplay between decentralization and classification. The first aspect appears through the distances between the nodes in the graph, while the second aspect plays through the moment generating functions of the likelihood ratios that describe the decision problem. By applying the derived closed-form relation to different network topologies and inference problems, we observe some interesting and perhaps unexpected behavior emerging. In particular, we characterize the scaling law (with the network size) for the loss over popular network topologies, showing that the error probabilities might differ by orders of magnitude; and we examine how performance is affected by the relative distance between informative and uninformative agents over the graph.

Abstract: Vision cues are available and informative for pedestrian action prediction, but obtaining stable target-centric representations from video frames remains challenging without frame-level external perception cues. Thus, most methods rely on additional perception modules or multi-source information fusion, leaving the reliability of vision-centric setting an open question. To this end, we propose ViCross, a vision-centric pedestrian crossing action prediction framework powered by multimodal large language models, which maintains target-centric reasoning from video frames without additional perception modules beyond first-frame target initialization. While multimodal large language models exhibit strong visual understanding, applying them directly to vision-centric action prediction faces two challenges. First, accurately perceiving target pedestrians often requires high resolution inputs and dense visual tokenization, making full-frame encoding computationally prohibitive. ViCross tackles this with Variable Resolution Patch Mapping module for efficient token allocation while preserving key pedestrian details. Second, missing spatiotemporal priors hinder consistent cross frame reasoning. ViCross mitigates this with a Spatial Constraint Enhancement Strategy that captures past motion, future locations, and action semantics for training-time proactive spatial rectification. Extensive experiments show that ViCross delivers clear gains in vision-centric prediction settings and is competitive with multi-source fusion approaches in several settings. Code is available at https://github.com/2tianyao1/ViCross.git.

Abstract: Local pedestrian-vehicle forecasting spans heterogeneous physical scales: pedestrians combine root locomotion with articulated motion, whereas vehicles are rigid bodies described by kinematic state and oriented extent. Existing road-agent forecasters typically omit pedestrian articulation, while pose forecasters leave vehicle futures outside the learned rollout. We introduce PV-WM, a history-only world model over structured post-perception tracks. It recurrently advances pedestrian root motion, 15-joint articulation, and learned vehicle states within a synchronized heterogeneous state. The generated pedestrian and vehicle chunks supply the next recurrent boundary; vehicle boxes are reconstructed from predicted center and heading with observed extent, and P-V geometry is recomputed after every transition. Relative to a matched one-shot complete-state predictor, recurrent execution reduces Root ADE by 12.7% and MPJPE by 14.8%. Feedback interventions show that later predictions depend on the content, temporal order, and pedestrian identity of generated articulation. Across 824 aligned Waymo contexts, with 797 providing valid future vehicle support, PV-WM reduces Root ADE by 5.2%, MPJPE by 7.6%, P-V distance error by 11.9%, and oriented-box closest-approach error by 5.8% relative to a validation-selected Modular Specialist. The single-network model uses 57.1% fewer parameters, 96.5% lower average FLOPs per local scene, and 25.5% lower measured p95 latency. PV-WM unifies this heterogeneous future state while preserving type-specific pedestrian and vehicle dynamics.

Abstract: Reliable risk assessment remains a central challenge for Autonomous Vehicles (AVs). Despite advances in automation, passenger cognition provides a non-intrusive auxiliary signal that improves both objective and perceived safety without requiring active human intervention. We introduce an Electroencephalogram (EEG)-based Brain-Computer Interface (BCI) that decodes passenger neural responses for both Risk Prediction (RP) and Danger Identification (DI), explicitly modeling humans as passengers to match real-world AV use. To achieve this, we propose the Passenger Cognitive Model (PCM), Risk-aware Sequential Labeling (RSL), and the Passenger EEG Decoding Strategy (PEDS), which integrates a 3D Convolutional Recurrent Neural Network (3D-CRNN) model for joint EEG decoding. Experimental results show that 3D-CRNN achieves a Balanced Accuracy (BA) of $95.3\% \pm 2.7\%$ in RP and improves single-subject DI from $80.9\% \pm 3.9\%$ to $85.0\% \pm 3.2\%$ with RSL. Event-wise analyses further show that 3D-CRNN consistently outperforms other models across different event types in RP and DI. In generalization experiments, 3D-CRNN achieves $77.0\% \pm 5.3\%$ BA in cross-session DI and $77.4\% \pm 1.1\%$ BA on seen subjects in cross-subject evaluation, while maintaining a $64.9\% \pm 8.5\%$ BA on unseen subjects, demonstrating promising generalizability and transferability across both intra-subject and inter-subject variability. These findings establish an Electroencephalogram (EEG) decoding framework for AV passenger hazard perception and suggest that passenger cognitive signals can provide auxiliary supervision for future AV decision-making and Safety of the Intended Functionality (SOTIF) support.

Abstract: In world model planning, sensing inputs pass through an encoder and predictor before affecting planner decisions, so final task success alone cannot reveal where sensing disturbances attenuate or persist in the pipeline. We apply 10 visual and temporal sensing degradations to a world model planner and track their effects across representation, future prediction, planner preference, and physical outcome using paired evaluation on the same 50 tasks. The relative impact of degradations was not preserved across stages: large representation shifts could attenuate downstream, while smaller initial shifts could persist to the outcome, and internal-response ordering did not directly match physical-outcome ordering. Temporal degradations also showed distinct patterns: even with similar overall changes in observation history, responses differed substantially with the location of corrupted information and the planner's actual exposure. This non-uniform stage-wise response was also observed in secondary evaluations with another manipulation task and a different world model. Stage-wise diagnosis can therefore identify where sensing disturbances attenuate or persist and help prioritize subsequent model verification and sensing mitigation.

Abstract: Visual-Inertial (VI) fusion is fundamental to accurate and robust state estimation, where camera and IMU measurements are combined according to their respective uncertainties. Existing methods, however, fuse the two modalities with predefined uncertainties, regardless of how reliable each is in the local context, and thus often struggle under challenging environments involving illumination changes, dynamic objects, and textureless regions. In this paper, we present MAC-I$^2$, which achieves robust VI fusion through learned metric-aware covariance for both modalities, so that vision and IMU compete on their own merits rather than relying on predefined uncertainties. Here, metrics-aware means that each predicted covariance faithfully reflects the actual magnitude of the corresponding measurement noise. On the visual side, we propagate learned feature-matching uncertainties into pose covariances for the fusion. On the inertial side, motivated by the observation that integration error accumulates sharply at the early stage and grows slowly afterward, we design a learned IMU model with a learnable initial covariance, and propose a dedicated fine-tuning strategy on a held-out training subset to enable the metrics-aware covariance on unseen sequences. As a showcase, we build a VI initialization and calibration system, since accurate and robust initialization and calibration are the prerequisite for any reliable VI system. Experiments on EuRoC, and VBR show that MAC-I$^2$ substantially outperforms existing methods: it achieves a 99.9% initialization success rate on EuRoC, reducing gravity and velocity errors by about 60% and 42% over the strongest baseline, and maintains 80% success rate on challenging VBR sequences where baseline methods such as VINS-Mono drop below 10%.

Abstract: Reasoning is a promising route to the generalization that autonomous driving requires in the long tail, as it can infer how the elements of a scene depend on one another and traverse those dependencies to conclusions beyond what is observed. Yet it is hard to tell whether a model's conclusions follow the scene's dependencies, because no driving representation makes them explicit enough to test against. Text-based reasoning traces lack spatio-temporal grounding, spatio-temporal scene graphs lack causal links, and reasoning annotations at scale are increasingly model-generated and hard to verify. To this end, we introduce CASCADE (Causal Spatio-Temporal Analysis of Driving Environments), which encompasses two components: (1) a structured scene representation for reasoning in driving scenes and (2) a human-annotated dataset built on it. For every actor that interacts with the ego vehicle, the CASCADE representation records frame-by-frame, for as long as the actor is visible, what action is taken, where it occurs, and how it depends on the actions and states of others. The resulting structure makes reasoning predictions machine-verifiable: they can be scored against it element by element, without relying on (M)LLM judges. The CASCADE dataset provides comprehensive human annotations for 2,066 driving clips of the PhysicalAI dataset, with over 34K elements that establish the spatio-temporal and causal context of each scene, including 8.6K time-stamped ego and agent actions, 3.7K causal links and 2.9K potential influences, and 6.1K annotations for agents, objects, traffic lights, and environments. Being entirely human-annotated, CASCADE provides the reference for this comparison: benchmarking the reasoning abilities of Physical AI models, and verifying the quality of automatically generated reasoning labels. The CASCADE dataset is available at https://huggingface.co/datasets/nvidia/cascade.

Abstract: 3D point cloud semantic segmentation is essential for real-world spatial understanding, yet the prohibitive cost of human annotations motivates unsupervised approaches that require no labels. Existing superpoint-based methods typically rely on spectral analysis at a fixed granularity, failing to capture the hierarchical semantic structures inherent in complex indoor scenes. To bridge this gap, we present a Multi-Scale Spatially-Constrained Partition (MSSP) framework that combines multi-scale spectral analysis with spatially-constrained clustering. Multi-scale spectral analysis constructs enriched superpoint descriptors across multiple clustering granularities; however, the resulting high-dimensional feature space calls for a structural prior to translate into cleaner segmentation. Spatially-constrained clustering supplies this prior by restricting superpoint merging to physically adjacent regions, imposing the spatial coherence needed for multi-scale features to be effective. Extensive experiments on S3DIS and ScanNet show that MSSP achieves the best mIoU among unsupervised methods on the main benchmarks, with particularly significant gains on S3DIS. Notably, our ablation reveals a regularize-then-enrich interaction: multi-scale features alone do not improve final segmentation, yet become highly effective when coupled with spatial regularization, underscoring that spatial coherence is aprerequisite for multi-scale representations in superpoint clustering.

Abstract: This letter investigates a reach-avoid game involving two Attackers and one Defender, where the Attackers aim to maximize the number reaching the target region while the Defender seeks to minimize it. In contrast to conventional complete information formulations, we consider an information asymmetry scenario where the Attackers' heterogeneous maximum speeds are privately known but publicly disclosed to lie within continuous ranges. Existing studies on uncertain speeds, however, have primarily focused on homogeneous settings, whereas heterogeneity extends the uncertainty from a common capability level to the relative capability configuration of the Attackers. To address the resulting capture-order ambiguity over infinitely many possible speed combinations, we establish a critical speed pair framework that characterizes when different capability configurations induce different optimal capture orders, and enables the analysis of the Defender's guessing behavior and the design of information-limiting strategies for the Attackers. We demonstrate that under certain initial conditions, the Attackers can mislead the Defender into making suboptimal decisions through a slow-speed deception strategy, achieving superior payoffs compared to the complete information game. Numerical visualizations reveal the widespread occurrence of such dilemma conditions.

Abstract: Synthesizing photorealistic driving videos along specified trajectories is essential for scalable closed-loop simulation. Reconstruction-based methods leverage neural rendering to synthesize geometrically consistent views, but often exhibit diverse artifacts and missing content when the viewpoint deviates from the training trajectory. In contrast, generative models can synthesize realistic views along arbitrary trajectories from vehicle sensor data, yet often struggle to maintain temporal and geometric consistency across frames. To combine the strengths of both, we propose PRG-Fusion, a framework for driving view synthesis that uses reconstruction evidence to orchestrate generative priors across regions. Specifically, we extract region-wise degradation evidence from reconstructed driving scenes and convert it into Preserve, Repair, and Generate (PRG) labels. At inference, these labels serve as a unified routing policy for region-aware spatiotemporal synthesis, orchestrating 3DGS appearance preservation, LiDAR-guided structural correction, and video-prior-driven content completion across Preserve, Repair, and Generate regions, respectively. We then follow a two-stage training paradigm, first establish geometric control from sparse LiDAR projections and subsequently learning appearance control from dense 3DGS renderings. Extensive experiments on Waymo demonstrate that PRG-Fusion achieves state-of-the-art overall performance in novel trajectory video synthesis, with superior visual quality and geometric fidelity while maintaining competitive view consistency under large trajectory shifts.

Abstract: Diffusion policies offer a powerful and expressive parameterization for continuous control. Yet, their integration with reinforcement learning remains conceptually and algorithmically challenging. In this work, we address this gap by introducing a noisy-space action-value (Q-)function that assigns values to diffusion latents through the distribution of executed actions induced by the denoising process. We show that this construction admits a precise semantic interpretation and derive a noisy-space policy gradient (NSPG) that optimizes noisy latents using only clean action-space value estimates. Building on this result, we formulate a KL-regularized policy improvement over noisy latents and show that the resulting objective admits a diffusion-compatible regression form, avoiding backpropagation through the denoising process. Empirical results on state-based D4RL benchmarks and vision-based OGBench tasks demonstrate that the proposed noisy-space objective provides a principled and effective basis for training diffusion policies in offline reinforcement learning. Project webpage: https://mahmoud-selim.github.io/NSPG/