In the context of autonomous driving and embodied AI technologies advancing into open-world scenarios, traditional 3D reconstruction methods are finding it difficult to meet real-time requirements in specific, enclosed environments. Meanwhile, conventional streaming 3D reconstruction techniques are grappling with a 'memory bottleneck.' To tackle these challenges, Alibaba's AutoNavi team has introduced ABot-Recon, the pioneering 10,000-frame-level streaming 3D reconstruction model that operates without long-range dependencies. This model utilizes a dual-engine framework consisting of 'local pose prediction + residual optimization.' By leveraging just 12 consecutive frames as a dynamic context window, it ensures a consistent level of computational complexity.

ABot-Recon showcases innovation in prediction mechanisms, error management, and resource utilization. According to testing results, it exhibits reduced trajectory errors, accelerated inference speeds, and a notable decrease in video memory usage—down to a mere one-third of that required by comparable mainstream models. This enables the model to run on consumer-grade graphics cards. Furthermore, the model achieves 10,000-frame-level 3D reconstruction using only monocular RGB video, eliminating the need for depth sensors or pre-calibrated camera parameters.

At present, the inference code and training weights for ABot-Recon have been made publicly available, along with a designated experience area for users to explore and interact with the technology.