Visuomotor policy achieves millimeter robotic docking with geometric features
GR-VIP: Geometric representation conditioning for precise visuomotor policy
Robotics
Summary
Getting robots to move precisely to a spot and angle is very hard, especially using cameras. The authors designed a new way for a robot to understand tiny differences between where it is and where it should be by using special visual features related to camera positions. Their robot could dock very accurately within millimeters and worked well even in new places it hadn't seen before or in low light. This method helps robots control their movements better by focusing on detailed visual clues instead of just general images.
What this means in practice
- •For mobile robot control teams: Implement precise docking maneuvers for robots in diverse real-world settings without retraining the control policy.
- •For industrial automation engineers: Enable robotic arms or mobile platforms to align parts or tools with millimeter accuracy using vision-based control.
Authors
Chanhwi Lee, Jiwon Seong, Eunje Choi, Hohyung Lee, Jongsoo Lee, Junyeong Bae, Inveom Kwak, Woosang Shin, Soohee Han
Abstract
Achieving precise terminal position and heading alignment remains a significant challenge for learning-based mobile robot control. Recent vision-based learning methods have shown promise for precise target-relative control and generalization to unseen environments with pretrained visual encoders, but general-purpose visual representations are not specifically designed to capture the fine-grained spatial differences between current and goal views required for precise control, and real-world demonstrations combining millimeter-scale terminal accuracy with zero-shot transfer across environments remain limited. To address these challenges, we propose Geometric Representation Conditioning for Precise Visuomotor Policy (GR-VIP), a goal-conditioned visuomotor policy that leverages a visual encoder pretrained for relative camera pose estimation. Rather than using the final relative-pose estimate, GR-VIP conditions action prediction on the dense relational features produced before pose regression, preserving fine-grained cross-view information for learning corrective actions from demonstrations. We evaluate GR-VIP on a real-world docking task and demonstrate high docking success rates and millimeter-scale terminal accuracy in both the training and unseen environments, without additional training for transfer. Furthermore, supplementary experiments illustrate the policy's ability to align with non-docking targets and complete docking under a tested low-light condition.