标题： ZipMPC：通过模仿学习的压缩上下文相关MPC成本 显示英文标题

标题： ZipMPC: Compressed Context-Dependent MPC Cost via Imitation Learning

摘要： 模型预测控制（MPC）的计算负担限制了其在实时系统（如机器人）中的应用，并且通常需要使用较短的预测范围。 这不仅影响控制性能，还增加了设计能够反映期望长期目标的MPC成本函数的难度。 本文提出了ZipMPC，一种通过为较短预测范围的MPC学习一个压缩且与上下文相关的成本函数来模仿长预测范围MPC行为的方法。 与替代方法（如近似显式MPC和自动成本参数调优）相比，ZipMPC在以下方面表现出更好的性能：i) 优化长期目标；ii) 保持与较短预测范围MPC相当的计算成本；iii) 确保约束满足；iv) 将控制行为推广到训练期间未观察到的环境。 为此，ZipMPC利用了带有神经网络的可微分MPC概念，以通过MPC优化传播模仿损失的梯度。 我们在自主竞速的仿真和真实世界实验中验证了所提出的方法。 ZipMPC始终比选定的基线更快完成圈数，达到接近长预测范围MPC基线的圈速。 在较短预测范围MPC基线无法完成一圈的挑战性场景中，ZipMPC能够完成。 特别是，这些性能提升也出现在训练期间未见过的赛道上。 显示英文摘要

摘要： The computational burden of model predictive control (MPC) limits its application on real-time systems, such as robots, and often requires the use of short prediction horizons. This not only affects the control performance, but also increases the difficulty of designing MPC cost functions that reflect the desired long-term objective. This paper proposes ZipMPC, a method that imitates a long-horizon MPC behaviour by learning a compressed and context-dependent cost function for a short-horizon MPC. It improves performance over alternative methods, such as approximate explicit MPC and automatic cost parameter tuning, in particular in terms of i) optimizing the long term objective; ii) maintaining computational costs comparable to a short-horizon MPC; iii) ensuring constraint satisfaction; and iv) generalizing control behaviour to environments not observed during training. For this purpose, ZipMPC leverages the concept of differentiable MPC with neural networks to propagate gradients of the imitation loss through the MPC optimization. We validate our proposed method in simulation and real-world experiments on autonomous racing. ZipMPC consistently completes laps faster than selected baselines, achieving lap times close to the long-horizon MPC baseline. In challenging scenarios where the short-horizon MPC baseline fails to complete a lap, ZipMPC is able to do so. In particular, these performance gains are also observed on tracks unseen during training.