Ziyan Gao

Achieving 86.79% action classification accuracy, this framework outperforms existing methods by over 80% in generating precise robotic commands from video demonstrations.

Instead of relying on hand-crafted heuristics, Con-DSO *learns* to identify and downweight unreliable visual data in RGB-D odometry, leading to substantial accuracy gains in challenging environments.

Robots can now learn manipulation skills from unstructured videos with significantly improved accuracy and generalization by decoupling video understanding from policy learning.

Forget hand-engineered pushing primitives: this unified policy uses visual prompts to achieve versatile and efficient object rearrangement.