The paper "Efficient Data Collection for Robotic Manipulation via Compositional Generalization" by Jensen Gao, Annie Xie, Ted Xiao, Chelsea Finn, and Dorsa Sadigh explores the effectiveness of data collection strategies in robotic manipulation to facilitate broad generalization. The authors investigate whether robot policies can compose environmental factors from their training data to succeed in unseen combinations, and how leveraging prior robotic datasets can enhance this compositional ability. Key findings include: - Robot policies exhibit compositional generalization, particularly when prior robotic datasets are used. - Data collection strategies that focus on capturing individual factor values and exploiting composition outperform those that naively optimize for coverage of all factor combinations. - Prior robotic datasets, such as BridgeData V2, are crucial for enhancing compositional generalization. - Policies trained using these strategies achieve a success rate of 77.5% when transferred to new environments with unseen combinations of environmental factors, compared to only 2.5% without prior data. The paper also discusses the importance of physical factors in compositional generalization and the challenges of generalizing to unseen factors. The authors propose efficient in-domain data collection strategies that exploit composition, which can reduce the amount of data needed while improving generalization.The paper "Efficient Data Collection for Robotic Manipulation via Compositional Generalization" by Jensen Gao, Annie Xie, Ted Xiao, Chelsea Finn, and Dorsa Sadigh explores the effectiveness of data collection strategies in robotic manipulation to facilitate broad generalization. The authors investigate whether robot policies can compose environmental factors from their training data to succeed in unseen combinations, and how leveraging prior robotic datasets can enhance this compositional ability. Key findings include: - Robot policies exhibit compositional generalization, particularly when prior robotic datasets are used. - Data collection strategies that focus on capturing individual factor values and exploiting composition outperform those that naively optimize for coverage of all factor combinations. - Prior robotic datasets, such as BridgeData V2, are crucial for enhancing compositional generalization. - Policies trained using these strategies achieve a success rate of 77.5% when transferred to new environments with unseen combinations of environmental factors, compared to only 2.5% without prior data. The paper also discusses the importance of physical factors in compositional generalization and the challenges of generalizing to unseen factors. The authors propose efficient in-domain data collection strategies that exploit composition, which can reduce the amount of data needed while improving generalization.