Reproduce “Evaluation of MPC-based Imitation Learning for Human-like Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Flavia Sofia Acerbo, Jan Swevers, Tinne Tuytelaars, Tong Duy Son
This paper evaluates the integration of imitation learning with differentiable Model Predictive Control (MPC) to achieve human-like autonomous driving. The work introduces MPC-based closed-loop state cloning, which enables robust learning against covariate shift and allows policies to capture human preferences such as comfort and lateral deviation patterns, verified in a high-fidelity driving simulator.
Key insights
- Differentiable MPC provides inherent stability for autonomous driving policies, and its parameters can implicitly learn and embed human driving preferences.
- Hierarchical policy decomposition, where a neural network influences MPC setpoints, enhances human-likeness and comfort, especially when controlling steering rate.
- Closed-loop state cloning (SC) is essential for robust imitation learning, effectively mitigating covariate shift and causal confusion in learned policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.