ASID: Active Exploration for System Identification in Robotic Manipulation This paper introduces ASID, a novel approach for real-world robotic control that combines autonomous exploration with system identification and policy optimization. ASID enables the learning of effective control strategies by using a small amount of real-world data to refine a simulation model and plan accurate control strategies that can be deployed in the real world. The key idea is to use an initial (possibly inaccurate) simulator to design effective exploration policies that, when deployed in the real world, collect high-quality data to refine the simulation model. ASID consists of three stages: exploration, system identification, and policy optimization. In the exploration phase, the goal is to collect informative data about unknown parameters by maximizing the Fisher information. This is achieved by training an exploration policy that maximizes the Fisher information, which quantifies the usefulness of the data collected. In the system identification phase, the collected data is used to refine the simulation model. Finally, in the policy optimization phase, a task-specific policy is trained in the updated simulator and then transferred to the real world. ASID has been evaluated on several robotic manipulation tasks, including sphere manipulation, laptop articulation, rod balancing, and shuffleboard. The results show that ASID can effectively identify unknown parameters of the real environment and learn policies that successfully transfer to the real world with only a small amount of real-world data. The approach is particularly effective in scenarios where the physical parameters of the environment are unknown and the task requires precise control. The paper also discusses the challenges of sim-to-real transfer and how ASID addresses these challenges by using a combination of exploration, system identification, and policy optimization. The approach is compared with several baselines and ablations, and the results show that ASID outperforms these methods in terms of performance and efficiency. The paper concludes that ASID provides a promising approach for real-world robotic control by combining autonomous exploration with system identification and policy optimization.ASID: Active Exploration for System Identification in Robotic Manipulation This paper introduces ASID, a novel approach for real-world robotic control that combines autonomous exploration with system identification and policy optimization. ASID enables the learning of effective control strategies by using a small amount of real-world data to refine a simulation model and plan accurate control strategies that can be deployed in the real world. The key idea is to use an initial (possibly inaccurate) simulator to design effective exploration policies that, when deployed in the real world, collect high-quality data to refine the simulation model. ASID consists of three stages: exploration, system identification, and policy optimization. In the exploration phase, the goal is to collect informative data about unknown parameters by maximizing the Fisher information. This is achieved by training an exploration policy that maximizes the Fisher information, which quantifies the usefulness of the data collected. In the system identification phase, the collected data is used to refine the simulation model. Finally, in the policy optimization phase, a task-specific policy is trained in the updated simulator and then transferred to the real world. ASID has been evaluated on several robotic manipulation tasks, including sphere manipulation, laptop articulation, rod balancing, and shuffleboard. The results show that ASID can effectively identify unknown parameters of the real environment and learn policies that successfully transfer to the real world with only a small amount of real-world data. The approach is particularly effective in scenarios where the physical parameters of the environment are unknown and the task requires precise control. The paper also discusses the challenges of sim-to-real transfer and how ASID addresses these challenges by using a combination of exploration, system identification, and policy optimization. The approach is compared with several baselines and ablations, and the results show that ASID outperforms these methods in terms of performance and efficiency. The paper concludes that ASID provides a promising approach for real-world robotic control by combining autonomous exploration with system identification and policy optimization.