Title: Seq-DeepIPC: Sequential Sensing for End-to-End Control in Legged Robot Navigation

Abstract:

This paper introduces Seq-DeepIPC, a novel end-to-end perception-to-control framework designed for legged robot navigation in real-world settings. By tightly coupling multi-modal perception—specifically RGB-D and GNSS data—with temporal fusion and control mechanisms, Seq-DeepIPC enhances intelligent sensing for autonomous legged locomotion. The architecture jointly performs semantic segmentation and depth estimation, thereby generating enriched spatial features that facilitate superior planning and control. To ensure efficient deployment on edge devices, the model employs a lightweight encoder, which significantly reduces computational load without compromising accuracy.

In a departure from traditional methods that rely on inertial measurement units (IMUs), our approach simplifies heading estimation by discarding noisy IMU data. Instead, global heading is derived through the differential analysis of sequential GNSS coordinates. We have curated a more extensive and diverse dataset comprising both road and grass terrains to support this work. The efficacy of Seq-DeepIPC was validated on a robot dog platform. Comparative and ablation studies demonstrate that while other baseline models do not benefit from sequential inputs, our approach sees marked improvements in both perception and control.

Seq-DeepIPC delivers competitive or superior performance relative to its model size. While the GNSS-only heading method shows reduced reliability in urban canyons with tall structures, it remains robust in open environments. Ultimately, Seq-DeepIPC extends the scope of end-to-end navigation from wheeled systems to more versatile, temporally-aware legged platforms. To facilitate further research, we will make our code available via GitHub at https://github.com/oskarnatan/Seq-DeepIPC.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC