The paper "Active Exploration for System Identification in Robotic Manipulation" by Memmel et al. proposes a novel approach called Active Exploration for System IDentification (ASEID) to address the challenges of model-free and model-based control strategies in robotic manipulation. The key idea is to use a small amount of real-world data to autonomously refine a simulation model and then plan an accurate control strategy that can be deployed in the real world. The approach relies on an initial (possibly inaccurate) simulator to design effective exploration policies, which, when deployed in the real world, collect high-quality data. The authors demonstrate the efficacy of this paradigm in identifying physical parameters such as articulation, mass, and other parameters in several challenging robotic manipulation tasks, showing that only a small amount of real-world data is needed for effective sim-to-real transfer. The paper is structured into several sections, including an introduction, related work, preliminaries, the proposed approach (AsID), experimental evaluation, and discussions. The introduction highlights the challenges of controlling robots in the real world using reinforcement learning and the limitations of model-based approaches. The related work section reviews system identification, simulation-to-reality transfer, and model-based reinforcement learning. The preliminaries section defines the decision-making setting as Markov Decision Processes (MDPs) and outlines the goal of learning a policy that maximizes reward in the true environment. The proposed approach, AsID, is a three-stage pipeline: exploration via Fisher information maximization, system identification, and solving the downstream task. The exploration phase aims to collect informative data by playing an exploration policy that maximizes Fisher information, which quantifies the usefulness of the data collected. The system identification phase updates the simulator parameters using the collected trajectory data. The downstream task phase trains a task-specific policy in the updated simulator and transfers it to the real world. The experimental evaluation section includes ablation studies and comparisons with baselines, demonstrating the effectiveness of the proposed approach in both simulated and real-world tasks. The authors evaluate the approach on tasks such as sphere manipulation, rod balancing, and shuffleboard, showing that ASID can accurately identify unknown parameters and successfully perform the downstream tasks with minimal real-world data.The paper "Active Exploration for System Identification in Robotic Manipulation" by Memmel et al. proposes a novel approach called Active Exploration for System IDentification (ASEID) to address the challenges of model-free and model-based control strategies in robotic manipulation. The key idea is to use a small amount of real-world data to autonomously refine a simulation model and then plan an accurate control strategy that can be deployed in the real world. The approach relies on an initial (possibly inaccurate) simulator to design effective exploration policies, which, when deployed in the real world, collect high-quality data. The authors demonstrate the efficacy of this paradigm in identifying physical parameters such as articulation, mass, and other parameters in several challenging robotic manipulation tasks, showing that only a small amount of real-world data is needed for effective sim-to-real transfer. The paper is structured into several sections, including an introduction, related work, preliminaries, the proposed approach (AsID), experimental evaluation, and discussions. The introduction highlights the challenges of controlling robots in the real world using reinforcement learning and the limitations of model-based approaches. The related work section reviews system identification, simulation-to-reality transfer, and model-based reinforcement learning. The preliminaries section defines the decision-making setting as Markov Decision Processes (MDPs) and outlines the goal of learning a policy that maximizes reward in the true environment. The proposed approach, AsID, is a three-stage pipeline: exploration via Fisher information maximization, system identification, and solving the downstream task. The exploration phase aims to collect informative data by playing an exploration policy that maximizes Fisher information, which quantifies the usefulness of the data collected. The system identification phase updates the simulator parameters using the collected trajectory data. The downstream task phase trains a task-specific policy in the updated simulator and transfers it to the real world. The experimental evaluation section includes ablation studies and comparisons with baselines, demonstrating the effectiveness of the proposed approach in both simulated and real-world tasks. The authors evaluate the approach on tasks such as sphere manipulation, rod balancing, and shuffleboard, showing that ASID can accurately identify unknown parameters and successfully perform the downstream tasks with minimal real-world data.