Presentation description

This project contributes to ongoing research aimed at enhancing robot interactions with novel objects.

Recent work in robot grasping has demonstrated that larger datasets lead to improved grasp performance and better generalization to novel objects, emphasizing the need for efficient, large-scale data collection. Additionally, collecting these data through realistic, gravity-enabled environments is important for developing optimal grasp models. To address these needs, we developed a software framework that efficiently produces large-scale datasets using parallelized, gravity-enabled simulations.

This software framework contributed to the proposal of a novel "First Plan, Then Evaluate" model for planning collision-free trajectories for the robot to reach and successfully grasp a target object. Initial results suggest that this method yields a higher likelihood of grasp success compared to traditional methods.

"First Plan, Then Evaluate" serves as the basis for continuing work on stable object placement at a target location. Traditional pipelines employ a "First Evaluate, Then Plan" model, which can be computationally expensive because it plans trajectories sequentially based on a ranked set of targets. Additionally, it is prone to discarding high-quality trajectories during the planning phase because they do not ideally reach the initial, pre-evaluated target. However, this novel approach plans all trajectories in parallel. The evaluator then ranks the resulting paths based on predicted success likelihood for all trajectories, and the robot executes the most promising option.

In addition to efficient and precise trajectory planning, stable object placement may require real-time adjustments to the object's orientation post-grasp. We will utilize a combination of visual and proprioceptive (joint angles) feedback to estimate and adjust post-grasp object orientation, ensuring a greater likelihood of stable placement.