Long-term non-prehensile planar manipulation is a challenging task for planning and control, requiring determination of both continuous and discrete contact configurations, such as contact points and modes. This leads to the non-convexity and hybridness of contact optimization. To overcome these difficulties, we propose a novel approach that incorporates human demonstrations into trajectory optimization. We show that our approach effectively handles the hybrid combinatorial nature of the problem, mitigates the issues with local minima present in current state-of-the-art solvers, and requires only a small number of demonstrations while delivering robust generalization performance. We validate our results in simulation and demonstrate its applicability on a pusher-slider system with a real Franka Emika robot.

翻译：长期非抓取式平面操作对于规划与控制而言是一项具有挑战性的任务，需要确定连续与离散的接触构型，例如接触点和接触模式。这导致了接触优化的非凸性与混合特性。为了克服这些困难，我们提出了一种将人类演示融入轨迹优化的新颖方法。实验表明，该方法能够有效处理问题的混合组合特性，缓解当前最先进求解器中存在的局部极小值问题，并且仅需少量演示即可实现稳健的泛化性能。我们在仿真环境中验证了结果，并在配备真实Franka Emika机器人的推-滑系统中展示了其适用性。