Unsupervised domain adaptation for vibration-based robotic ground classification in dynamic environments

Vibration-based Robotic Ground Classification (V-RGC) is an ability of a field robot identifying the ground types (e.g., grass and clay) according to the proprioceptive vibration sequence. It has proved that the accurate and read-time V-RGC contributes a lot to the avoidance of non-geometric hazards, route planning, and pose estimation. However, V-RGC in a dynamic environment (i.e., data distribution may drift) has not yet been taken into consideration, which motivates us to propose a novel classification method named Joint Domain Adapta-tion Semi-Supervised Extreme Learning Machine (JDA-S2ELM). First, the projected maximum mean discrepancy (MMD) criterion is introduced to expand the classification boundaries from the source domain to the target domain in a computationally-efficient way. Second, the joint-distribution domain adaptation (DA) is proposed to realize a cascaded marginal -and conditional-distribution DA training framework, which shows a higher and more stable accuracy. Third, target-domain manifold regularization is added to smooth the classification boundaries to cut through low-density regions, thus further increasing the target-domain classification accuracy. The real-world experiment demonstrates that the proposed JDA-S2ELM could increase the target-domain accuracy from about 30% to 90%, which means that V-RGC is adaptable to a dynamic environment.