GUILD: Guided Iterative Data Collection for Imitation Learning
GUILD uses estimates of which robot demonstrations improve performance to guide what gets collected next.
Vision-language models propose environments, object configurations and goals for data collection. The system then estimates how collected trajectories affect policy performance in the target deployment setting and feeds that assessment into the next collection round. In simulation and real-robot tests, the authors report better sample efficiency and deployment performance than passive collection, several baselines and experienced roboticists. The abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
Vision-language models propose environments, object configurations and goals for data collection. The system then estimates how collected trajectories affect policy performance in the target deployment setting and feeds that assessment into the next collection round. In simulation and real-robot tests, the authors report better sample efficiency and deployment performance than passive collection, several baselines and experienced roboticists. The abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
score 4