Combining Coarse and Fine Physics for Manipulation Using Parallel-in-Time Integration

We present a method for fast and accurate physics-based predictions during non-prehensile manipulation planning control. Given an initial state sequence of controls, the problem predicting resulting states is key component variety model-based control algorithms. propose combining coarse (i.e. computationally cheap but not very accurate) predictive physics model, with fine expensive to generate hybrid model that at required speed accuracy given task. Our approach based on Parareal algorithm, parallel-in-time integration used computing numerical solutions general systems ordinary differential equations. adapt combine pushing off-the-shelf engine deliver are as run in substantially less wall-clock time, thanks parallelization across time. use these model-predictive-control framework trajectory optimization, plan actions avoid obstacle reach goal location. show models, we can achieve same success rates planner uses directly, significantly faster. experiments simulation real robotic setup. Videos available here: https://youtu.be/5e9oTeu4JOU .