Driving · 2026-10-23
TerraTransfer++: Demonstration-Free End-to-End Planner
Zikang Xiong, Weixin Li, Zhouchonghao Wu, Akshay Rangesh, Saarth Bonde, Grantland Hall, Chen Tang, Yihan Hu, Wei Zhan
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
TerraTransfer++ trains a self-play driving policy in a vectorized simulator, then distills its rollouts into a flow-matching trajectory teacher that also outputs plans. Finally, it aligns a vision student (frozen head, trainable adapters) to reproduce the teacher via action KL divergence, a batch-relational low-rank structural loss, and (for planning) trajectory-matching. The authors report demonstration-free training with no logged-trajectory supervision at any stage.
Why it matters
The authors propose a demonstration-free recipe for end-to-end autonomous driving that avoids the data scarcity and alignment mismatch of imitation learning, and the rendering/per-step cost of closed-loop RL in photorealistic simulation. By reusing a self-play teacher for both closed-loop driving and camera-conditioned planning, they claim improved benchmark performance without human trajectory demonstrations.
Method
- Train a single self-play policy in TerraZero using PPO in a multi-agent vectorized simulator; the policy learns driving competence without imitating human trajectories.
- Freeze the self-play teacher and distill its self-play futures into a transformer trajectory teacher with a flow-matching trajectory head; targets are realized futures of the self-play teacher, not logged humans.
- Align a camera-conditioned vision student to either frozen teacher using (i) action KL divergence, (ii) a batch-relational low-rank structural loss on scene-feature relations, and (iii) for the planner, a trajectory-matching loss.
Limitation
The self-play teacher’s ability is limited by scenario realism: even the privileged teacher cannot make progress on many Extreme scenarios, consistent with visually implausible initial configurations (e.g., other cars passing through each other).
Abstract (from arXiv)
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.
Related papers
- RoXDrive: Closed-Loop Reinforcement Learning for End-to-End Autonomous Driving via Action-Faithful Rollouts
- SHRED: Self-Play for Heterogeneous Road-Users without Expert Demonstrations
- YOUDrive: Driving Style as a Composable Direction for End-to-End Autonomous Driving
- AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving
- A Survey on End-to-End Autonomous Driving Training from the Perspectives of Data, Strategy, and Platform
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-21 · Indexed by Pub-AI: 2026-10-23 UTC
TerraTransfer++: Demonstration-Free End-to-End Planner
Preprint posted: 2026-10-21 · Original record ↗
- arxiv
- 2610.30319
Explicit source versions (1)
- v1 · 2026-10-21
- Preprint available
Source date: 2026-10-21 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.