UMI-Underwater: Learning Underwater Manipulation without Underwater Teleoperation
Summary: arXiv:2603.27012v1 Announce Type: cross
Underwater robotic manipulation presents a unique set of challenges, primarily due to the unpredictable nature of the underwater environment. The visibility in such conditions is often compromised by factors such as lighting variations and water turbidity, making it difficult for robots to perform grasping tasks effectively. In response to these challenges, researchers have developed a groundbreaking system named UMI-Underwater that showcases the potential of autonomous learning in underwater environments.
Key Features of UMI-Underwater
- Autonomous Data Collection: The UMI-Underwater system is designed to autonomously collect successful underwater grasp demonstrations. This self-supervised data collection pipeline enables the system to gather a diverse array of underwater interaction data without the need for extensive human intervention.
- Knowledge Transfer: One of the most significant innovations of this system is its ability to transfer grasp knowledge from on-land human demonstrations to underwater scenarios. This is achieved through a depth-based affordance representation, which effectively bridges the gap between on-land and underwater environments.
- Robustness to Environmental Changes: The affordance model, which is trained using on-land handheld demonstrations, is deployed in underwater settings without the need for additional training. This zero-shot deployment is made possible through geometric alignment techniques that ensure the model remains effective despite variations in lighting and color.
- Control Action Generation: To facilitate effective manipulation underwater, an affordance-conditioned diffusion policy is trained on the collected underwater demonstrations. This training allows the system to generate precise control actions tailored for underwater environments.
Experimental Results
In extensive pool experiments, UMI-Underwater demonstrated significant improvements in grasping performance compared to traditional methods. The system exhibited enhanced robustness to background shifts and displayed a remarkable ability to generalize its grasping capabilities to objects that were only seen during on-land training. These advancements are particularly noteworthy when compared to RGB-only baselines, which often struggle with the complexities of underwater manipulation.
Conclusion
The development of UMI-Underwater marks a significant step forward in the field of underwater robotics. By leveraging self-supervised learning and effective knowledge transfer techniques, this system not only simplifies the process of underwater manipulation but also broadens the scope of potential applications in marine research, underwater exploration, and environmental monitoring.
For more information, including access to the code, videos, and additional results, please visit the UMI-Underwater project page at https://umi-under-water.github.io.