Overview
Domain randomisation works, but how you schedule it matters more than how much of it you apply. This paper measures that effect across 4,000 real grasp attempts.
Journal of Field Robotics
2025
10.1002/rob.22187
We study how randomisation schedules affect the transfer gap in vision-based grasping policies, and propose a curriculum that reduces real-world failure rates by 27% without any real-world fine-tuning.
Domain randomisation works, but how you schedule it matters more than how much of it you apply. This paper measures that effect across 4,000 real grasp attempts.
A tightly-coupled LiDAR-inertial odometry pipeline for sub-$500 sensor stacks, with 0.38% drift over 12-hour warehouse runs.
Treating battery state as a first-class planning constraint extends inspection coverage by 19%.