Deliberate Practice: Budget-Optimal Robot Skill Learning

An innovative algorithm known as Deliberate Practice (DP) allows robots to learn skills independently within a constrained practice budget for sequential tasks. This method, outlined in a paper submitted to arXiv, determines a budget-optimal allocation by focusing on skills that yield the highest expected cumulative reward while remaining learnable within the set budget. DP assesses the time required to master skills and the cumulative rewards associated with the task plans that these skills enable. A significant aspect of this research is a bilinear program that accurately computes the optimal allocation using standard solvers. Experiments conducted in both simulated and real-world long-horizon manipulation tasks illustrate that this approach enables robots to effectively utilize limited practice time to develop beneficial policies and enhance long-horizon planning. The paper, categorized under Computer Science > Robotics, was submitted on August 26, 2025 (arXiv:2608.13415). This research tackles the complex issue of managing numerous skill plans within a large practice budget, demonstrating the algorithm's practicality through various simulations and physical robot tests. The findings significantly advance robot learning by offering a structured method for time allocation, essential for real-world applications where resources are scarce.