日本フィジカルAI新聞

世界のフィジカルAIを、日本語で。

週刊ニュースレター購読
自動運転/自己対戦/蒸留arXiv:2610.30319

TerraTransfer++:専門家デモ不要のエンドツーエンド自動運転プランナー

TerraTransfer++: Demonstration-Free End-to-End Planner

シェア:XThreadsFacebookLINEはてブBluesky

ベクトル化シミュレータでの自己対戦によりポリシーを事前学習し、そのロールアウトをフローマッチング軌道教師へ蒸留、凍結視覚バックボーン上の軽量アダプタで再現することで、専門家デモなしにエンドツーエンド自動運転プランナーを学習する手法。

詳しい要約

1. どんなもの?

自己プレイ由来の知識だけで学習する、人間の expert demonstration を必要としない end-to-end 自動運転の学習フレームワーク。vectorized simulator での self-play により単一の reactive な policy を pretrain し、その rollout を flow-matching trajectory teacher に distill した上で、frozen な vision backbone に lightweight adapter を付けて 2 種類の end-to-end student(closed-loop policy と planner)を得る。

2. 先行研究と比べてどこがすごい?

- imitation learning は fleet-scale の driving log を要し、安全クリティカルな tail の被覆が薄い - photorealistic simulator での closed-loop RL は sample-efficiency、per-step の rendering コスト、大規模 vision backbone の forward pass がボトルネック - 提案法は vectorized simulator の self-play(毎秒数百万 step)で多様な state 分布を得て、logged trajectory への supervised alignment を一切行わない点が異なる - 結果として HUGSim で公開済みの imitation ベース end-to-end 手法を上回り、NAVSIM navhard で 59.4 EPDMS(DriveZero-Scale の 57.1 超)、navtest で 95.5 PDMS を human demonstration なしで達成

3. 技術・手法の肝は?

- vectorized simulator での self-play により単一 policy を pretrain(基本的な driving-quality reward のみ) - その rollout を flow-matching trajectory teacher に distill(teacher 自身も planning を行う) - frozen な pretrained vision backbone 上に lightweight adapter を学習し、teacher の latent space を再現 - 損失は action KL divergence、batch-relational low-rank structural loss、planner 用の trajectory-matching loss - 全ての target が self-play policy に由来し、logged trajectory に対する supervision を行わない - 学習データは (image, scene-state) の paired frame のみで…

4. どうやって有効だと検証した?

- photorealistic な 3D Gaussian splatting ベースの closed-loop benchmark HUGSim で、公開済みの imitation ベース end-to-end 手法すべてを上回る closed-loop policy を確認 - NAVSIM の navhard で 59.4 EPDMS(DriveZero-Scale の 57.1 を上回り公開手法中最高)、navtest で 95.5 PDMS を human demonstration なしで達成

5. 議論はある?

- 要旨からは不明(self-play の reward 設計の感度、sim-to-real gap、adapter 容量や backbone 凍結の影響、計算コストの内訳などは明示されていない)

6. 次に読むべき論文は?

- DriveZero-Scale(NAVSIM navhard での比較対象) - HUGSim(closed-loop benchmark) - NAVSIM(navhard / navtest の評価基盤) - imitation learning ベースの end-to-end 自動運転手法、closed-loop RL、flow-matching による trajectory 生成、self-play を用いた driving policy 学習の関連研究

※ AIが要旨から生成した要約です。正確性は原文をご確認ください。

著者: Zikang Xiong, Weixin Li, Zhouchonghao Wu, Akshay Rangesh, Saarth Bonde, Grantland Hall, Chen Tang, Yihan Hu, Wei Zhan

分類: cs.RO, cs.AI, cs.CV, cs.LG

原文アブストラクト

End-to-end autonomous driving has achieved state-of-the-art performance on benchmarks and real-world deployments. Its standard training recipe, however, is expensive across all existing paradigms: imitation learning requires fleet-scale driving logs whose behavior distribution only thinly covers the safety-critical tail; closed-loop reinforcement learning in a photorealistic simulator, while offering diverse behaviors, is bottlenecked by sample-efficiency, per-step cost of rendering, and a forward pass through a large vision backbone. Self-play in vectorized simulators changes the economics: at millions of rollout steps per second, one can obtain a state distribution with natural diversity in collisions, near-misses, and recoveries that no driving log contains, and a highly reactive policy emerges from a basic driving-quality reward. Our approach exploits this asymmetry. We pretrain a single policy by self-play, distill its rollouts into a flow-matching trajectory teacher that also plans, and train lightweight adapters on a frozen pretrained vision backbone to reproduce the latent space of either teacher through the action KL divergence, a batch-relational low-rank structural loss, and, for the planner, a trajectory-matching loss. Every target traces back to the self-play policy, so alignment never supervises against a logged trajectory: a paired dataset of (image, scene-state) frames suffices, with no need for the curated expert demonstrations that imitation pretraining is built on. From this one self-play teacher we obtain two end-to-end students: a closed-loop policy that outperforms every published imitation-based end-to-end method on the photorealistic 3D Gaussian splatting closed-loop benchmark HUGSim, and a planner that reaches 59.4 EPDMS on NAVSIM navhard, above every published method including DriveZero-Scale (57.1), and 95.5 PDMS on navtest, without human demonstrations.

PR本紙発行元 EmplifAI