GUILD preprint reports more efficient collection of robot training data

Researchers report in a GUILD preprint that guided collection of robot demonstrations improved sample efficiency and deployment performance in simulation and real robot tests, compared with passive collection strategies, multiple baselines and experienced roboticists. The method alternates between proposing tasks, collecting demonstrations and evaluating how those demonstrations affect a policy's performance in its target deployment setting. Vision-language models suggest relevant environments, object configurations and goals; estimates of each trajectory's contribution then guide subsequent collection. The approach addresses the cost of collecting demonstrations through remote human control. Its evaluation step accounts for a specific limitation: semantically relevant tasks do not necessarily improve a robot policy.