ZipDrift: One-Step Drift Generation for Multimodal MPC Cost Correction

Ke Zhang proposes ZipDrift, which uses a one-step drift generator to learn cost corrections for short-horizon MPC; training relies only on cost evaluations of candidate corrections and does not backpropagate through the optimizer. In a kinematic overtaking benchmark simulation based on ZipMPC, 53% of scenarios are bimodal, and handcrafted short-horizon MPC collides in 28.5% of test episodes. With eight-dimensional per-stage corrections, ZipDrift achieves zero collisions across all 600 test episodes and matches the best evaluated candidate; as a control, an idealized ZipMPC-style regressor has a collision rate of 4.0% (16.0% when expert labels are inconsistent), and random candidates ranked by the critic under the same budget have a collision rate of 2.0%. These results come from a preliminary simulation study and are not real-vehicle validation.

AI 根据报道生成 · 6 days ago更新