-Integrate the existing capabilities of multiple algorithms within Haomo, develop a comprehensive foundational largemodel solution, and ultimately apply the capabilities of this foundational large model to practical business scenarios.-Address real-world logistics vehicle perception challenges and deliver an advanced logistics vehicle perception solution.

-Utilized the data closed-loop system to perform data analysis at a scale of over a hundred million, resulting in significant cost savings in labor and achieving a weekly automatic release.-Data-Centric AI: Drawing inspiration from solutions like MoCo and DINO for unsupervised data training. Pre-trained models show an improvement of over 10% in comparison to on-vehicle models. Active learning techniques are employed to annotate high-value scenarios from extensive datasets. Finally, joint training is conducted with data from other domains, unsupervised, and actively learned data to enhance downstream task performance.-Multi-Modality: Leveraging the capabilities of MLLM, autonomous driving scenario understanding is achieved by combining image and language spaces. Scenario retrieval and captioning are performed through techniques related to CLIP/BLIP.-AIGC: Data synthesis is carried out using techniques such as GAN and Stable Diffusion to supplement rare and dangerous scenarios, achieving performance up to 72% of real data. Instruct Pix2Pix and LoRA are utilized to train SD models, with ControlNet constraining semantics, annotations, and other information.

-Developed, deployed an end to end data driven prediction module in self-driving cars.

-Was in charge of roadside perception of the AVP project, led a team up to 10 people. Defined technical architect, product form, and participated in the daily operation, successfully passed the test for the very first time.-Defined the quantitative test method for the modules, found out the boundary of the current form.

-Independently initiated field perception for the automated valet parking project. -Finished data collection, data annotation, camera calibration, algorithms, and hardware selection in one-person. Designed the first version of the software architecture and the form of the product.-Using cameras to segment, classify and localize the object.-Using traditional and deep learning technics to implement a computer vision product.

-Designing environment model algorithms for the downstream groups with fault handling.-Integrating multiple sources not limited to high definition map, GPS, and IMU.-Constructing local map which gives the autonomous vehicle better perception.-Supporting high way and valet parking autonomous driving.