velocity-aware sampling

This work addresses the challenge of correspondence ambiguity in video frame interpolation caused by large-scale nonlinear motion and complex occlusions. To this end, the authors propose a dynamic motion trajectory–based feature scanning mechanism that constructs feature sequences along nonlinear paths guided by optical flow. The method introduces a learnable residual velocity update and a velocity-aware state space model (SSM) to enable adaptive dense sampling and feature aggregation in fast-moving regions. Integrated with an end-to-end jointly optimized intermediate flow estimation and occlusion-aware refinement module, the proposed approach achieves state-of-the-art performance on standard benchmarks, particularly excelling in scenarios involving large displacements and intricate dynamic content.

This work addresses the computational inefficiency of motion planning for high-degree-of-freedom robots operating under dynamic constraints in complex environments. The authors propose a high-speed trajectory planning method grounded in differential flatness, which maps the dynamics into a flat output space to enable analytical, time-parameterized trajectory generation. By integrating SIMD-based parallel acceleration with a sampling-based planning framework, the approach achieves, for the first time, a general-purpose, highly accurate, and ultra-fast trajectory planner for differentially flat systems—including robotic arms, ground vehicles, and aerial robots. Experiments demonstrate that the method generates dynamically feasible trajectories in microseconds to milliseconds in both simulated and real-world cluttered dynamic environments, substantially improving planning efficiency while rigorously preserving dynamic feasibility and tracking accuracy.

This work proposes an end-to-end adaptive sampling and denoising framework tailored for real-time path tracing under extremely low sampling budgets—typically below one sample per pixel. Conventional super-resolution methods struggle with spatially varying noise, reconstruction difficulty, and perceptual importance, often leading to detail loss. To address this, the framework employs a differentiable stochastic sampling strategy that enables gradient estimation through discrete sampling decisions, coupled with perceptually driven tone-mapped training to avoid oversampling in visually insensitive regions. It further integrates a pyramid-aggregation denoising filter and a learnable albedo demodulation module. Experiments demonstrate that the method significantly outperforms uniform sampling, with particularly notable improvements in perceptually critical areas such as specular highlights and shadow boundaries.

Sampling-based motion planners suffer from degraded performance in high-dimensional configuration spaces—particularly those containing narrow passages—due to low sampling efficiency. To address this, we propose an online adaptive multi-resolution sampling framework that dynamically balances sparse and dense sampling without requiring prior training or hand-crafted heuristics. Our key innovation is a selective refinement mechanism enabling real-time adjustment of sampling granularity. By integrating uniform random sampling across multi-resolution spatial exploration with a bidirectional search strategy, the algorithm prioritizes sparse exploration in SE(2), SE(3), and ℝ¹⁴ configuration spaces. Experimental results demonstrate that our approach significantly outperforms state-of-the-art sampling-based planners on complex terrains and the Franka Emika Panda robotic platform, achieving superior computational efficiency, probabilistic completeness, and practical deployability.

This work addresses two key challenges: efficient sampling from unnormalized densities and reward-based fine-tuning of generative models. We propose Tilt Matching, a novel algorithm that reinterprets the flow matching velocity field as the sum of a stochastic interpolation term and the cumulant of a reward function. By formulating a reward-tilted stochastic optimal control problem via a dynamical equation, our method implicitly solves for the tilted velocity field—without requiring explicit reward gradients or trajectory backpropagation. Innovatively, we introduce a covariance-driven velocity correction and a reward-tilting mechanism, substantially reducing the variance of the objective. Empirically, Tilt Matching achieves state-of-the-art performance on Lennard-Jones potential sampling and matches top-tier methods in reward fine-tuning of Stable Diffusion—while eliminating the need for heuristic reward scaling and naturally supporting few-step flow mapping architectures.

This study quantifies the amplification effect of early perturbations on subsequent steps in diffusion model sampling. For the first time, this work directly compares “shaped” noise derived from historical trajectories with newly sampled “fresh” noise, systematically revealing their differential impacts on generative updates through paired angular gain analysis, latent-space RMS measurements, and perceptual feature evaluations. The results demonstrate that shaped noise yields angular gain ratios ranging from 1.14 to 2.35, with such gains being highly dependent on trajectory alignment. By elucidating how structured noise propagates differently than uncorrelated noise throughout the denoising process, this research provides a novel perspective for understanding error propagation mechanisms in diffusion sampling. Furthermore, it establishes a theoretical foundation for optimizing sampling strategies in generative modeling.

Existing sampling-based motion planners struggle to simultaneously achieve real-time performance, near-optimality, and parallel scalability under complex dynamic constraints. This work proposes a novel parallelized kinodynamic sampling-based planner that decouples the traditionally sequential planning pipeline into three massively parallel subroutines, constructs a sparse trajectory tree, and focuses propagation and optimization on high-potential nodes within local neighborhoods. The method is the first to provide asymptotic δ-robust near-optimality guarantees within a large-scale parallel framework, overcoming the fundamental limitation of conventional parallel SBMP approaches that cannot optimize objective functions. Experimental results demonstrate up to three orders of magnitude faster solution times compared to state-of-the-art serial methods, while also yielding higher-quality solutions than the best existing GPU-based planners.

This work addresses the challenge of efficiently allocating integration timesteps for flow-matching models under a fixed computational budget to enhance generation quality. The authors propose SharpEuler, a training-free sampler that, for the first time, identifies trajectory acceleration as the dominant source of Euler discretization error. Leveraging this insight, they offline estimate a sharpness profile of the pretrained velocity field and derive a sharpness-based power-law timestep density. By combining finite-difference estimation, profile smoothing, and quantile transformation, SharpEuler constructs a solver-aware timestep grid that enables stable and efficient Euler integration at any inference budget. Experiments demonstrate that SharpEuler significantly improves sample quality, reduces mode leakage, and enhances multimodal coverage.

This work addresses the resampling bias in uniform sampling on implicit manifolds caused by errors in local density estimation. To mitigate this issue, the authors propose an adaptive hybrid density estimation method built upon the MASEM framework. By analyzing the distribution of inter-shell distances among k-nearest neighbors, a gating mechanism is introduced: when these distances significantly deviate from an exponential distribution, a high-order moment estimator based on Polynomial Moment Maximization (PMM2/PMM3) is activated; otherwise, the method defaults to plug-in or maximum likelihood estimation (MLE). This adaptive strategy reduces mean squared error in density estimation by 22%–36% under asymmetric gamma and boundary-spacing scenarios, while explicitly delineating its regime of validity to achieve a balance between accuracy and robustness.

This work addresses the high computational cost and slow inference speed of video generation strategies in robotic tasks by proposing a training-free inference paradigm for diffusion models. The method introduces, for the first time, a draft-and-target sampling mechanism into video generation, integrating self-play denoising, token chunking, and a progressive acceptance strategy to simultaneously generate global trajectories and verify fine-grained details within a single model. This parallelized approach substantially reduces redundant computation. Experimental results demonstrate that the proposed method achieves up to a 2.1× speedup on three robotic task benchmarks while maintaining near-identical task success rates, significantly enhancing inference efficiency without compromising performance.