<p>Existing RGB and Thermal video object detection (RGBT VOD) models typically rely on cross-modal aligned data, while alignment process is usually manual and thus time-consuming and labor-intensive. Cross-modal alignment methods are often adopted to mitigate the spatial differences of unaligned data, but would suffer from poor generalization due to the intrinsic modality gap between RGB and thermal modalities in challenging scenarios. To this end, this paper presents a novel ImageBind guided Progressive Transformation Network called IPTNet, which leverages the generalization ability of pretrained multimodal foundation model to adapt diverse challenging cross-modal alignment scenarios, for alignment-free RGBT VOD. In particular, we design an aligner with a corresponding alignment consistency all built upon the pretrained multimodal foundation model. The aligner adopts a progressive two-stage alignment strategy, consisting of a linear affine transformation module for coarse locating and a warping-based affine module for fine-grained refinement. Each stage is guided by a dedicated alignment consistency loss to ensure progressively constrained and accurate alignment in the multimodal feature space. In addition, to fully exploit the benefits of multimodal fusion under severe modality receptive field discrepancy, we design a Thermal-to-RGB fusion module to enable high-fidelity feature integration in a specific region of thermal modality. Our proposed IPTNet shows superior performance compared with existing RGBT VOD methods on the challenging UVT-VOD2024, VT-VOD50, CVC-14 and LLVIP benchmark datasets. Especially on the UVT-VOD2024 dataset, our method surpasses its benchmark by approximately 13%.</p>