RoboGene researchers report improved robot learning from generated tasks

Researchers report that robot models trained on manipulation tasks generated by RoboGene achieved higher success rates and stronger generalization in real-world experiments. Their arXiv paper describes collecting 18k trajectories using a framework that proposes tasks for single-arm, dual-arm and mobile robots. RoboGene samples tasks for diversity, assesses their quality through automated self-reflection and uses a memory of human-verified feedback to refine proposals. The approach addresses the physical cost of collecting robot interaction data by broadening coverage across objects, scenes and skills. Its validation process still incorporates human feedback.