Researchers Jingxu Gu and Dongze Lian of Nanjing University report an overall score of 85.55% for DEED, their method for fine-grained robot manipulation, on the five-task MetaFine benchmark. Their preprint was submitted to arXiv on Oct. 29.
The paper compares that score with 43.01% for the published π0 baseline, a benchmark result reported by the authors.
DEED, short for Demonstration Expansion and Execution Design, adapts a pretrained vision-language-action policy. The authors say it pairs different instructions with the same initial scene, re-renders trajectories with different camera and lighting conditions while keeping their states and action targets, and uses hierarchical sampling to balance the training data.
For execution, the method converts predicted action chunks into robot commands through inverse kinematics, task-specific gripper mappings and closed-loop replanning. Its five benchmark tasks cover part-level grasping, directional motion, switch toggling, object sorting and letter insertion.