https://wangyida.github.io/tags/policy-fine-tuning/ Recent content in Policy Fine-Tuning on Yida Wang https://wangyida.github.io/logos/android-chrome-512x512.png https://wangyida.github.io/logos/android-chrome-512x512.png Hugo en Fri, 05 Jun 2026 11:31:20 +0000 https://wangyida.github.io/posts/craft/ Fri, 05 Jun 2026 11:31:20 +0000 https://wangyida.github.io/posts/craft/ <blockquote> <p>Re-direct to the full <a href="https://arxiv.org/abs/2605.04470"><strong>PAPER</strong></a> and <a href="https://currychen77.github.io/CRAFT/"><strong>PROJECT PAGE</strong></a></p> </blockquote> <p>CRAFT is a reinforcement fine-tuning framework for autonomous driving policies that combines dense counterfactual supervision with grounded closed-loop residual feedback. It treats counterfactual trajectory scoring as a broad proxy signal, then uses executed on-policy rollouts to correct interaction-dependent failures, while an EMA teacher preserves reliable pre-trained behavior during adaptation. Across multiple driving-policy families, this design improves driving score and success rate in closed-loop Bench2Drive evaluation.</p>