Tsukada Lab. Computer Network Group
← Publications Int'l conference

To Route or to Adapt? Multi-Expert Diffusion Planners in Unseen Regions

Yun Li, Ehsan Javanmardi, Simon Thompson, Jiachen Gong, Qunli Zhang, Zifan Zeng, Shiming Liu, Peng Wang, Yidu Zhang, Manabu Tsukada

Conference on Robot Learning (CoRL 2026), Austin, Texas, USA, Nov 9, 2026

Summary in three lines

  • The paper tests whether selecting among trained diffusion-planning experts can transfer driving to unseen regions.
  • Across three Aichi maps, eight routing methods failed to beat a static geographically matched expert on the reported fleet metric.
  • Adapting with 20 target-region scenes reduced trajectory error, and replay limited forgetting on the original domains.

Figures from the paper

Figure 1: Trajectory generation and expert selection
Figure 1. Trajectory generation and expert selectionSource: Figure 1 from Li et al., Conference on Robot Learning (CoRL 2026) (2026) (authors’ arXiv version 2610.31095, CC BY 4.0)The diagram separates routing among pre-existing trajectory candidates from adapting the generator itself.

Abstract

Diffusion trajectory planners with domain-specific experts face a practical question in unseen regions: can they rely on selecting among existing experts, or must the planner itself adapt? We study this question using a planner with a shared adapter and per-domain expert adapters. Across three real-world held-out maps, none of eight test-time routing schemes improves the fleet-mean lateral offset over the static default, which uses the same expert for every scene. Even an oracle that selects the best expert for each scene reduces the mean offset by only 0.9 m. In contrast, reward-based fine-tuning on just 20 target-region scenes yields substantially larger gains. On matched evaluation slices, joint replay achieves a fleet-mean offset of 6.9 m, compared with 12.4 m for the static default. Comparisons under the same target-region training budget show that target-region training drives the improvement, while replaying training scenes limits forgetting on the original ODDs. On nuPlan, fine-tuning across training cities also improves open-loop offsets, although closed-loop performance deteriorates. Further diagnostics reveal that two test-time adaptation baselines achieve low offsets by producing degenerate trajectories. These results show the limited headroom for selecting among existing experts and the greater open-loop gains available from adapting the planner.

Code

Related research and software

How to cite

Y. Li, E. Javanmardi, S. Thompson, J. Gong, Q. Zhang, Z. Zeng, S. Liu, P. Wang, Y. Zhang, and M. Tsukada, “To Route or to Adapt? Multi-Expert Diffusion Planners in Unseen Regions,” in Conference on Robot Learning (CoRL 2026), Austin, Texas, USA, 2026.

BibTeX

@inproceedings{Li2026e,
  title         = {To Route or to Adapt? Multi-Expert Diffusion Planners in Unseen Regions},
  author        = {Yun Li and Ehsan Javanmardi and Simon Thompson and Jiachen Gong and Qunli Zhang and Zifan Zeng and Shiming Liu and Peng Wang and Yidu Zhang and Manabu Tsukada},
  url           = {https://arxiv.org/abs/2610.31095 https://github.com/tlab-wide/RouteOrAdapt},
  year          = {2026},
  date          = {2026-11-09},
  urldate       = {2026-11-09},
  booktitle     = {Conference on Robot Learning (CoRL 2026)},
  address       = {Austin, Texas, USA},
  abstract      = {Diffusion trajectory planners with domain-specific experts face a practical question in unseen regions: can they rely on selecting among existing experts, or must the planner itself adapt? We study this question using a planner with a shared adapter and per-domain expert adapters. Across three real-world held-out maps, none of eight test-time routing schemes improves the fleet-mean lateral offset over the static default, which uses the same expert for every scene. Even an oracle that selects the best expert for each scene reduces the mean offset by only 0.9 m. In contrast, reward-based fine-tuning on just 20 target-region scenes yields substantially larger gains. On matched evaluation slices, joint replay achieves a fleet-mean offset of 6.9 m, compared with 12.4 m for the static default. Comparisons under the same target-region training budget show that target-region training drives the improvement, while replaying training scenes limits forgetting on the original ODDs. On nuPlan, fine-tuning across training cities also improves open-loop offsets, although closed-loop performance deteriorates. Further diagnostics reveal that two test-time adaptation baselines achieve low offsets by producing degenerate trajectories. These results show the limited headroom for selecting among existing experts and the greater open-loop gains available from adapting the planner.},
  keywords      = {CRONOS},
  pubstate      = {published},
  tppubtype     = {inproceedings}
}