Manipulation task planning for heterogeneous object stacking based on reinforcement learning

• المؤلف الرئيسي: Pham, Minh Khang
• مؤلفون آخرون: Domenico Campolo
• التنسيق: Final Year Project
• اللغة: English
• منشور في: 2019
• الموضوعات: DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics; DRNTU::Engineering::Mechanical engineering::Robots
• الوصول للمادة أونلاين: http://hdl.handle.net/10356/76393
• المؤسسة: Nanyang Technological University

The paper propose a new way of solving the Pallet Packing Problem by modelling it as a Markov decision process. This allows the program to make decisions step-by-step based only on the current state and adapt to any error in execution. By applying reinforcement learning techniques, an agent can be trained from simulation to learn a model-free near-optimal policy that maximize the discounted cumulative rewards, which is proportional to the original objective function. Experiments show positive results on simulations involving packing up to 12 boxes into a grid-based pallet of size 8*8*6.