Salvato in:
Dettagli Bibliografici
Autori principali: Lu, Hengtong, Huang, Victor Shea-Jay, Yang, Chengmin, Jing, Pengfei, Dai, Jifeng, Xie, Yan, Zhu, Benjin
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2605.12625
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866910220938117120
author Lu, Hengtong
Huang, Victor Shea-Jay
Yang, Chengmin
Jing, Pengfei
Dai, Jifeng
Xie, Yan
Zhu, Benjin
author_facet Lu, Hengtong
Huang, Victor Shea-Jay
Yang, Chengmin
Jing, Pengfei
Dai, Jifeng
Xie, Yan
Zhu, Benjin
contents Continuous-action policies trained on a single demonstrated trajectory per scene suffer from mode collapse: samples cluster around the demonstrated maneuver and the policy cannot represent semantically distinct alternatives. Under preference-based evaluation, this caps best-of-N performance -- even oracle selection cannot recover what the sampling distribution does not contain. We introduce DIAL, a two-stage Driving-Intent-Amplified reinforcement Learning framework for preference-aligned continuous-action driving policies. In the first stage, DIAL conditions the flow-matching action head on a discrete intent label with classifier-free guidance (CFG), which expands the sampling distribution along distinct maneuver modes and breaks single-demonstration mode collapse. In the second stage, DIAL carries this expanded distribution into preference RL through multi-intent GRPO, which spans all intent classes within every preference group and prevents fine-tuning from re-collapsing around the currently preferred mode. Instantiated for end-to-end driving with eight rule-derived intents and evaluated on WOD-E2E: competitive Vision-to-Action (VA) and Vision-Language-Action (VLA) Supervised Finetuning (SFT) baselines plateau below the human-driven demonstration at best-of-128, with the strongest prior (RAP) capping at Rater Feedback Score (RFS) 8.5 even with best-of-64; intent-CFG sampling lifts this ceiling to RFS 9.14 at best-of-128, surpassing both the prior best (RAP 8.5) and the human-driven demonstration (8.13) for the first time; and multi-intent GRPO improves held-out RFS from 7.681 to 8.211, while every single-intent baseline peaks lower and degrades by training end. These results suggest that the bottleneck of preference RL on continuous-action policies trained from demonstrations is not only how to update the policy, but to expand and preserve the sampling distribution being optimized.
format Preprint
id arxiv_https___arxiv_org_abs_2605_12625
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Driving Intents Amplify Planning-Oriented Reinforcement Learning
Lu, Hengtong
Huang, Victor Shea-Jay
Yang, Chengmin
Jing, Pengfei
Dai, Jifeng
Xie, Yan
Zhu, Benjin
Robotics
Computer Vision and Pattern Recognition
Continuous-action policies trained on a single demonstrated trajectory per scene suffer from mode collapse: samples cluster around the demonstrated maneuver and the policy cannot represent semantically distinct alternatives. Under preference-based evaluation, this caps best-of-N performance -- even oracle selection cannot recover what the sampling distribution does not contain. We introduce DIAL, a two-stage Driving-Intent-Amplified reinforcement Learning framework for preference-aligned continuous-action driving policies. In the first stage, DIAL conditions the flow-matching action head on a discrete intent label with classifier-free guidance (CFG), which expands the sampling distribution along distinct maneuver modes and breaks single-demonstration mode collapse. In the second stage, DIAL carries this expanded distribution into preference RL through multi-intent GRPO, which spans all intent classes within every preference group and prevents fine-tuning from re-collapsing around the currently preferred mode. Instantiated for end-to-end driving with eight rule-derived intents and evaluated on WOD-E2E: competitive Vision-to-Action (VA) and Vision-Language-Action (VLA) Supervised Finetuning (SFT) baselines plateau below the human-driven demonstration at best-of-128, with the strongest prior (RAP) capping at Rater Feedback Score (RFS) 8.5 even with best-of-64; intent-CFG sampling lifts this ceiling to RFS 9.14 at best-of-128, surpassing both the prior best (RAP 8.5) and the human-driven demonstration (8.13) for the first time; and multi-intent GRPO improves held-out RFS from 7.681 to 8.211, while every single-intent baseline peaks lower and degrades by training end. These results suggest that the bottleneck of preference RL on continuous-action policies trained from demonstrations is not only how to update the policy, but to expand and preserve the sampling distribution being optimized.
title Driving Intents Amplify Planning-Oriented Reinforcement Learning
topic Robotics
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2605.12625