Imitation learning (IL) shows its superiority in faster strategy optimization by leveraging expert demonstrations. However, as one of the powerful IL methods, behavior cloning (BC) suffers from covariate shift problems, where the agent’s policy drifts away from the expert’s, leading to compounding errors and decreased generalization performance. To this end, we propose Search-based Inverse Dynamics Imitation Learning, namely SIDIL, to enhance the robustness of imitation learning by augmenting expert demonstrations via trajectory perturbation and stitching. Specifically, SIDIL first employs a nearest-neighbor search method to find the closest points between expert and perturbed data, generating new actions near the expert data via a stitching strategy. By exploiting this, the agent can recover from deviations and complete tasks under a wider range of conditions. Experimental results on various robotic tasks show that SIDIL outperforms baseline algorithms with higher success rates across multiple tasks. Meanwhile, SIDIL is allowed to expand the attractive region around expert demonstrations by stitching states from expert demonstrations, enjoying a higher task completion success rate even for states outside the expert distribution. These results highlight our potential to enhance humanoid locomotion and dexterous robotic hand manipulation, making it particularly suitable for industrial manufacturing automation where adaptability, robustness, and safety are essential.

Offline imitation learning (IL) enables embodied agents of AI-driven automation control systems to acquire policies from demonstrations without interacting with the environment. However, collecting extensive expert demonstrations that rely on a manually designed reward function is labor-intensive, and even impractical for dynamic real-world settings. To mitigate the reliance on large amounts of reward-annotated expert data, previous works have made great efforts by leveraging both expert and suboptimal demonstrations. In this paper, we reveal that existing methods still suffer from suboptimal results for two reasons: (i) reward signals in suboptimal demonstrations are not reliable, due to these rewards derived from inexperienced annotators, thus can not reflect the actual rewards, and (ii) those suboptimal demonstrations, which are significantly distant from expert behaviors, result in low reconstructed rewards. Moreover, motivated by the recent diffusion-based foundation model, we propose Expert-Like Transition Imitation Learning (ELTIL), a novel generative trajectory augmentation method that effectively leverages diffusion models. ELTIL re-designs the reward function based on expert proximity, and then proposes amplified return conditioning for achieving high-quality trajectory generation. More specifically, instead of relying on environment-provided rewards or manually designed rewards, ELTIL first serves as a good reward labeler for transitions by measuring their distance to expert states. These re-designed rewards then serve as guidance in the diffusion process, with amplified returns contributing as conditioning values to bias generation toward expert-consistent behaviors. Extensive experiments on diverse D4RL benchmarks demonstrate that ELTIL consistently improves the performance and stability of reward learning and offline IL methods, maintaining reliability under varying noise levels, reward sparsity, and dataset distributions. These characteristics make ELTIL well-suited for deployment in robust and reliable automated control systems such as robotic manipulation and autonomous driving.

Offline reinforcement learning (RL) demonstrated remarkable performance in learning valid policies by benefiting from high-quality offline datasets. However, collecting such a dataset is labor-intensive, especially for humanoid locomotion. For this reason, many data augmentation techniques have been proposed to improve the quality of offline datasets through noise injection or data synthesis. However, existing data augmentation methods are noise-sensitive, resulting in limited capability in complex robotic environments. To address these issues, we propose guided amplified learning with Lipschitz constraint (GALC), a novel trajectory augmentation method that employs the reward-amplification-guided conditional diffusion model for noise-insensitive data augmentation. Specifically, we introduce a local Lipschitz continuity constraint to regulate the reverse denoising process from the offline dataset. Consequently, the exploration of the diffusion model can be restricted within the local continuity region of the original dataset, thereby generating high-reward trajectories. Moreover, the generated trajectories are also enforced to be noise-insensitive to perturbations, thus enjoying robustness. Notably, our proposed method can prevent the generation of unsafe actions that do not align with the environment dynamics. Extensive experiments on sparse reward scenarios and high-dimensional robotic tasks show that our proposed GALC achieves significant improvements in both the augmented trajectories and policy performance.

Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intra-class variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors.