From Grasping to Skills
Six real-robot trials exploring how grasping procedures and verification skills carry across sessions.
How can a robot reuse what it learns across separate interactions with an AI agent? This case study follows six successive attempts by Astra and a real XLeRobot to grasp an empty milk carton and hold it clear of a desk. Each attempt starts in a fresh workspace and conversation; reviewed procedures and reusable code carry experience forward, while model weights remain unchanged.
Four attempts succeeded and two failed. The failures exposed weaknesses in grasp enclosure, stability checks, and recovery, informing the procedures used in later attempts. Among successful runs, execution time decreased from 31 minutes 39 seconds to 18 minutes 45 seconds. This is an exploratory sequence with human interventions, rather than a controlled demonstration of consistent improvement.
The repository includes the full report, per-attempt records, control programs, skill updates, video previews, and measurement definitions.