Conceptos Básicos
Our closed-loop pipeline integrates interactive perception with online refined axis estimation, enabling adaptive and precise control during manipulation tasks involving articulated objects.
Resumen
The paper presents a novel approach for manipulating articulated objects, which combines interactive perception techniques with online axis estimation derived from SAM2-based tracking of 3D point clouds.
Key highlights:
- Closed-loop integration of interactive perception and real-time axis estimation: By continuously refining the motion axis based on updates from 3D point clouds, the approach enables adaptive and precise control during manipulation tasks, overcoming the limitations of open-loop methods.
- Utilization of advanced segmentation models for precise articulation recognition: The pipeline employs Grounding DINO for object detection and SAM2 for accurate segmentation, allowing the robot to isolate moving components and calculate the motion axis of articulated objects in real-time.
- Enhanced generalization and precision in articulated object manipulation: The method significantly improves manipulation precision and generalization across diverse cabinets in door-opening and drawer-opening tasks, outperforming state-of-the-art baselines and providing consistent axis-aware manipulation.
The experiments demonstrate that the proposed approach outperforms baseline methods, especially in tasks that demand precise axis-based control, such as opening doors to wider angles or drawers to greater extents.
Estadísticas
The robot needs to open the door larger than 8.6°, 10°, 20°, 30°, 40°, 45°, 50°, 55°, 60°, 65°, 70°.
The robot needs to open the drawer larger than 10 cm, 15 cm, 20 cm, 25 cm, 30 cm, 35 cm, 40 cm, 45 cm.
Citas
"Our closed-loop integration of interactive perception with online refined axis estimation allows for more precise and adaptive control, addressing the limitations of traditional methods that operate in an open-loop fashion."
"By continuously updating the robot's understanding of the object's kinematic state, our approach ensures that the robot maintains effective control throughout the manipulation task."