CeRLP: A Cross-embodiment Robot Local Planning Framework for Visual Navigation

A research team led by Haoyu Xi has introduced CeRLP, a novel framework designed to solve a major bottleneck in robotics: visual navigation across different robot designs. Traditional methods require collecting massive datasets for each new robot configuration or time-consuming model fine-tuning, often failing to account for physical robot geometry. CeRLP bypasses this by abstracting visual information into a unified geometric formulation, making it applicable to heterogeneous robots with varying dimensions, camera parameters, and camera types.

The core of CeRLP is a two-part innovation. First, it employs a depth estimation scale correction method that uses offline pre-calibration to resolve the scale ambiguity inherent in monocular depth estimation, recovering precise metric depth images. Second, it features a visual-to-scan abstraction module that projects diverse visual inputs into height-adaptive laser scans. This creates a robot-agnostic input for the navigation policy, making it robust to hardware differences. Experiments in simulation and extensive real-world tests on tasks like point-to-point and vision-language navigation demonstrate CeRLP's superior obstacle avoidance and generalization capabilities compared to previous methods.