Research Focus
Autonomous Robot Control Balancing Task Execution and Collision Avoidance
The research ensures a robot can successfully complete its main tasks while safely working unattended. We achieve this by training the robot within a self-play reinforcement learning framework, allowing it to continuously improve both task execution and collision avoidance. To build deep spatial awareness, the system distills multimodal vision data and internal motor proprioception into a single model that detects potential crashes. If the robot is about to hit itself or an unseen object, this real-time policy anticipates the move to prevent self-destruction and minimize damage to the surroundings. Ultimately, this approach enables reliable, autonomous robot operation in unpredictable environments without the risk of working hazzard.
Efficiency Optimization of Object Grasping by Robot Arm Pushing
This study integrates pushing manipulation by a robot arm using visual prompts. When humans organize objects on a desk, they not only push an object to a target position but also use different manipulation skills depending on the situation. For example, they gather scattered objects into a single area (grouping) or separate a single object from a cluttered pile (singulation). This research develops a flow-matching-based generative model that integrates these three skills, pushing, grouping, and singulation, into a single policy. In addition, a large vision-language model is used to interpret visual prompts given as points on an image and to understand the intended manipulation. Through this approach, the robot can select and execute appropriate actions from visual information, even in complex real-world environments. Ultimately, the study aims to build an autonomous system capable of performing flexible, advanced object manipulation in a human-like manner.