Robotics

ARC-Calib is a fully autonomous, model-based markerless camera-to-robot calibration framework that leverages exploratory robot motions and geometric optimization to achieve robust, generalizable calibration across diverse robots and scenarios, without requiring markers, pre-trained models, or extensive data collection.

rt-RISeg is a real-time, model-free framework for robot interactive segmentation of unseen objects. By leveraging robot interactions and analyzing body frame-invariant features, it can segment objects without any learned model. The method achieves significantly higher accuracy than previous approaches and can also enhance the performance of vision foundation models by providing high-quality segmentation masks.

We propose a novel concept called “Caging in Time” that enables robust object manipulation with a single end-effector under uncertainties and limited perception. This approach allows caging configurations to be formed in time, demonstrating effective open-loop manipulation without requiring detailed object knowledge or real-time feedback.

We propose a collision-inclusive planning framework for occluded object grasping that allows and exploits intentional collisions through compliant robot motions, demonstrating effective manipulation in scenarios where traditional collision-free approaches are insufficient.

The 2024 RGMC at ICRA featured five tracks testing autonomous robotic manipulation capabilities, drawing participation from teams worldwide to address challenges in manufacturing and essential manipulation skills.

We propose a robust, learning-free robotic peg-in-hole assembly system that leverages compliant, contact-rich interactions to eliminate both perception and execution uncertainties. By actively seeking contact and exploiting environmental constraints, our method enables reliable assembly across diverse scenarios without precise perception or prior training.