PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.31095v1cs.RO

TL;DR

In unseen regions, it is unclear whether a planner with domain-specific experts can rely on selecting among them or needs to adapt. The paper compares routing with adapter mixing and fine-tuning, finding little routing headroom but larger open-loop gains from target-region adaptation. Closed-loop performance remains a limitation.

  • Problem

    For planners with domain-specific experts, the paper asks how much selecting among existing experts can achieve in unseen regions compared with updating the planner.

    Source: §1 · p. 1

  • Method

    The study compares routing among expert trajectories with geographic weight mixing and fine-tuning of a shared-plus-expert diffusion planner.

    Source: §1 · p. 1, §5.1 · p. 6

  • Results

    6.9 m versus 12.4 m: joint replay achieves the lower fleet-mean offset on matched evaluation slices, while none of eight routing schemes improves over the static default across three held-out maps.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The results favour updating the planner over selecting among existing experts to reduce open-loop lateral offset in unseen regions.

    Source: §7 · p. 9

    Fine-tuned models score below the unadapted planner in nuPlan’s reactive closed-loop simulator in all four held-out cities, and joint-replay adapters have not been evaluated on a physical vehicle.

    Source: §7 · p. 9

Abstract

from arXiv · show

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.

1 Introduction

The paper asks whether unseen-region planning should select among existing domain experts or adapt the planner, which can risk degrading performance on training domains. It finds little benefit from routing and larger open-loop gains from target-region fine-tuning, while closed-loop performance remains a concern.

1 Introduction

  • The study compares selection among existing expert trajectories with methods that modify adapter weights in a shared-plus-expert planner.The planner has a shared adapter and four per-ODD expert adapters, evaluated on three held-out maps.Source: §1 · p. 1, Figure 1 · p. 2
  • None of eight routing schemes improves fleet-mean lateral offset over the static default across the three held-out maps; oracle selection gains only 0.9 m.The default uses the geographically nearest expert, ODD-Nagoya, on all three maps.Source: §1 · pp. 1–2
  • On matched evaluation slices, joint replay achieves a fleet-mean offset of 6.9 m versus 12.4 m for the static default.Comparisons indicate target-region training drives the gain, while replay primarily preserves performance on original ODDs.Source: §1 · p. 2
  • On nuPlan, fine-tuning across training cities improves open-loop offsets but degrades closed-loop performance.Two test-time adaptation baselines also achieve low offsets with nearly stationary or scene-independent trajectories.Source: §1 · p. 2

2 Related Work

The related work situates this study among trajectory uncertainty methods, geographic conditioning, expert routing and weight merging, and multi-task adaptation. The paper’s focus is selection or weight mixing of already-trained domain experts, alongside fine-tuning approaches.

Diffusion planning and conformal prediction

  • Prior trajectory and semantic uncertainty work includes conformal prediction, while selective conformal risk control combines abstention with calibrated prediction.This paper examines whether a conformal corridor can select among existing expert trajectories under deployment-region shift.Source: §2 · p. 2

Geographic conditioning in driving

  • GeCo and Drive-Anywhere train for regional or unseen-environment generalization, whereas this study guides selection or weight mixing of already-trained experts without backbone retraining.GeCo uses geographic channel attention and contrastive imitation learning; Drive-Anywhere uses multimodal foundation-model representations.Source: §2 · p. 2

Expert routing and weight merging

  • Related work spans online expert selection, mixtures of denoising experts, CLIP-based LoRA retrieval, and model soups or weight-space merging.Here, geographic mixing weights depend on deployment location.Source: §2 · p. 3

Multi-task learning and domain adaptation

  • The study relates shared-plus-expert adapters to multi-task learning and tests perturbed-scene fine-tuning, target-region training with replay, and test-time adaptation.Replay is used to limit forgetting; degenerate predictions from test-time adaptation are analyzed separately.Source: §2 · p. 3

3 Method

The method separates trajectory generation from candidate selection, then compares eight routing schemes with three mechanisms that alter or mix adapter weights. Evaluation uses lateral offset across defined held-out-map pools, with a safety gate applied after the reported selector output.

Generator–Selector decomposition

  • A diffusion Generator combines a shared LoRA with four per-ODD experts to emit four candidate trajectories, and a Selector ranks them and returns one.The architecture follows the Autoware Generator–Selector framework.Source: §3 · p. 3

Eight routing schemes

  • Eight routing schemes choose or score existing candidates using scene features, lane geometry, trajectory consistency, uncertainty, denoising loss, or reward.The schemes include a scene-feature MLP, lane-following, self-consistency variants, a split-conformal corridor, denoising-loss scoring, a reward-supervised gate, and reward-as-scorer.Source: §3 · p. 3

Three constructive mechanisms

  • Three constructive mechanisms mix geographic expert weights, fine-tune only the shared adapter on perturbed training scenes, or fine-tune the full stack with target-region and training scenes.Geographic mixing does not optimise parameters; joint replay uses target-region scenes together with 200 training scenes.Source: §3 · p. 3

Reward and safety gate

  • A downstream Safety Gate rejects trajectories violating hard limits, but reported lateral offsets are measured on Selector output before that gate.The inherited trajectory reward includes forward progress, route alignment, comfort, and safety terms.Source: §3 · p. 4

Evaluation protocol

  • The static default selects the geographically nearest training ODD expert, and oracle headroom measures the per-scene gain from choosing the lowest-offset candidate.Oracle headroom does not bound gains from weight mixing or fine-tuning.Source: §3 · p. 4
  • Pool A contains 24 routing scenes, while Pool B uses evaluation scenes from lanelets disjoint from each seed’s training scenes; results from the pools are reported separately.The lateral-offset metric is maximum absolute lateral displacement relative to the initial lanelet axis, and the fleet mean weights the three held-out maps equally.Source: §3 · p. 4

4 Limited Headroom for Expert Routing

Across the held-out maps, routing does not improve fleet-mean offset over the static default, and even oracle selection has little headroom. Diagnostics show that the tested routing schemes and representations do not reliably identify better experts under cross-region shift.

4 Limited Headroom for Expert Routing

  • None of the eight routing schemes improves the 13.74 m static-default fleet mean; six are significantly worse (p ≤0.012).The static default uses ODD-Nagoya, the best fixed expert on all three held-out maps; the conformal corridor’s mixed map-level results leave its fleet mean higher than the default.Source: Table 2 · p. 5, §4 · p. 5, Figure 3 · p. 6
  • Oracle selection improves mean offset by only 0.9 m on Pool A, with gains of 1.3 m on a 104-scene pool and 1.1 m on Pool B.Joint replay reduces Pool A mean offset by 9.1 m; on 17 scenes outside its training set, the improvement is 12.7 m (p = 0.0008).Source: §4 · p. 5
  • The scene-conditioned MLP assigns most held-out scenes to ODD-Hongo, self-consistency selects the shared module alone, and denoising loss poorly tracks realised offset.The planner reward often selects ODD-Shalun; the reward-supervised gate largely recovers the static default, while the conformal corridor either abstains or reduces to geographic mixing at the stated thresholds.Source: §4 · pp. 5–6
  • Leave-one-out classifiers do not beat the majority-class baseline, and the conformal predictor’s training and held-out residuals fail an exchangeability test (p = 0.0035).The tested representations and routing methods therefore do not reliably identify better experts under cross-region shift.Source: §4 · p. 6

5 Beyond Expert Selection: Weight Mixing and Adaptation

Across the adaptation alternatives, joint replay with target-region scenes yields the strongest supported fleet-mean result, while geographic weight mixing does not improve on the static default. The section also shows that low lateral offset can accompany degenerate trajectories.

5.1 Geographic Weight Mixing

  • At τ = 200 km, geographic weight mixing has a 15.86 m fleet mean versus 13.74 m for the static default, and neither mixing approach improves the fleet mean.The τ = 200 km mixture slightly improves Knowledge Hub and Ishiodai but increases Morikoro’s offset from 31.9 to 38.9 m.Source: §5.1 · p. 7

5.2 Shared-Adapter Repair with Geometric Perturbations

  • The geometric-perturbation repair trains for 500 steps on original-ODD scenes with lane rotations, centreline warps, and random lane-boundary removal.The three perturbations introduce variations in lane orientation, curvature, and boundary structure relevant to the open-park map.Source: §5.2 · p. 7
  • With the ODD-Kashiwa expert, shared-adapter repair reaches 11.1 m on Morikoro versus 31.9 m for the static default, while improving the other two maps.Lanelet rotation contributes the largest improvement in the three-seed ablations.Source: §5.2 · p. 7

5.3 Joint Replay with Target-Region Scenes

  • 6.91 ± 2.16 m on Pool B versus 12.4 m for the static default; joint replay transfers gains from Ishiodai training scenes to the other held-out maps.Across eight seeds, the sample-complexity results also show reduced seed variation as the target training set grows to 20 scenes.Source: §5.3 · p. 7
  • Under matched training budgets, target-only methods achieve approximately 4–8 m offsets, while joint replay limits summed training-ODD regression to 0.9 ± 1.4 m at 6.91 ± 2.16 m target offset.Target-only L2-SP has the lowest target offset, 4.04 m, but 95th-percentile jerk is 27% higher than joint replay; replay mainly preserves original-ODD performance.Source: §5.3 · p. 7

5.4 Variance-Collapse Diagnostics for Test-Time Adaptation

  • TENT reaches 3.37 m offset while its cross-scene trajectory distance falls from over 100 m to approximately 10 m; the hypernetwork produces only 12 m of forward progress over 8 s.The target-region scenes contain no neighbouring agents, so collision rates and time-to-collision cannot be evaluated.Source: §5.3 · p. 7, §5.4 · p. 8

6 Cross-City Evaluation and Real-Vehicle Runs

Cross-city fine-tuning lowers open-loop lateral offset on nuPlan, but the reported physical-vehicle runs concern the base planner rather than the joint-replay adapters.

6 Cross-City Evaluation and Real-Vehicle Runs

  • Fine-tuning on three training cities reduces Singapore open-loop lateral offset from 8.64 to 5.60 ± 0.78 m, a 35% reduction, with improvements also observed in the other held-out cities.The four leave-one-city-out splits use no data from the held-out city.Source: §6 · p. 8
  • The reward-trained shared-plus-expert base planner completed one ODD-Kashiwa vehicle run with a 47.7% comfort-violation rate versus 37.7% for a human driver, while a second run stopped after 72 m.The joint-replay adapters have not been tested on the vehicle.Source: §6 · p. 9

7 Discussion, Limitations, and Conclusion

The discussion favors adapting the planner over selecting existing experts for reducing open-loop offset, while emphasizing boundaries in closed-loop performance and evaluation scale. It also warns that low-offset test-time adaptation can reflect trajectory collapse.

7 Discussion, Limitations, and Conclusion

  • Selecting among the four experts offers limited improvement even with per-scene oracle selection, whereas small amounts of target-region data produce larger open-loop offset reductions.Under matched conditions, replay mainly helps retain performance on original ODDs rather than accounting for target-region gains.Source: §7 · p. 9
  • The static default’s success on the primary maps appears to reflect ODD-Nagoya’s strength rather than proximity alone.In a suburban area near ODD-Kashiwa, the nearest expert is 1.4 m worse than ODD-Nagoya; in a dense urban region nearest ODD-Hongo, that expert performs worst.Source: §7 · p. 9
  • Held-out-map results are open-loop; in nuPlan, fine-tuned models score below the unadapted planner in all four held-out cities, with significant differences in three.For Singapore, simulator score falls from 0.858 to 0.598 and drivable-area compliance from 0.97 to 0.76; joint-replay adapters have not been evaluated on a physical vehicle.Source: §7 · p. 9
  • The primary held-out evaluation covers only 4–10 road lanelets per map, and the full eight-method routing comparison remains on the smaller pool.The findings concern the tested experts, routing methods, and representations; applicability to monolithic planners trained on pooled data remains unknown.Source: §7 · p. 9

Overview of the Appendix

A.1 TENT Recipes and Variance-Collapse Diagnostics

  • TENT-online’s best fleet mean, 4.11 m at epoch 1, comes from variance collapse rather than learning the new region.Across ten epochs its K-sample variance loss falls from 40.6 to 0.0095 while fleet mean rises to 5.09 m; TENT from the static default reaches 3.37 m but cross-scene L2 collapses to 10.2 m.Source: §A.1 · p. 13

B.1 Recipe, Results and Diagnostic Probe

The hypernetwork maps deployment location to LayerNorm offsets, but its low lateral-offset results reflect nearly stationary trajectories rather than a constructive adaptation mechanism.

B.1 Recipe, Results and Diagnostic Probe

  • The hypernetwork maps the deployment location’s ECEF unit vector to offsets for the decoder’s 36,608 LayerNorm affine parameters.It is trained against recorded ego trajectories on paired training-ODD scenes, then conditioned at deployment on the target-map centroid.Source: §B.1 · p. 14
  • The best single-seed fleet mean is 5.74 m, while reducing hidden width from 64 to 16 raises the fleet mean to 11.0 m.The three-seed result is 6.27 ± 0.59 m.Source: §B.1 · p. 14
  • The diagnostic probe finds 12 m of forward progress at 1.5 m/s, making the low lateral offset an artifact of nearly stationary trajectories.The paper retracts hypernetwork-LN as a constructive mechanism and retains it as a documented failure mode.Source: §B.1 · pp. 14–15

C Derivation and Weights of the vMF Mixture

The vMF mixture interprets location-based expert weights as a posterior over training ODDs, interpolating between nearest-expert selection and uniform weighting. Weight averaging and merging do not reliably beat the best single expert on the held-out evaluation pool.

C Derivation and Weights of the vMF Mixture

  • At τ→0 the mixture becomes the static default, and at τ→∞ it becomes a uniform expert average.Intermediate temperatures interpolate between these cases.Source: C Derivation and Weights of the vMF… · p. 15, Table 10 · p. 16
  • The weights are a Bayesian posterior over training ODDs, obtained by applying a softmax to negative great-circle distance from the deployment location to each ODD centre.With a uniform prior and a likelihood that decays exponentially with distance, Bayes’ rule gives p(D = d | ℓ) = αd(ℓ).Source: C Derivation and Weights of the vMF… · p. 15
  • At τ=200 km, four geographic distance metrics yield Morikoro mixture weights agreeing to four decimal places, with largest L1 deviation 7×10−5.The compared metrics are Haversine, ECEF chord, equirectangular flat-Earth approximation, and local tangent plane (UTM).Source: C Derivation and Weights of the vMF… · p. 15, Table 9 · p. 16

D Weight-Averaging Baselines: Learned Coefficients, Model Soups and TIES-Merging

  • On the seed-7 evaluation pool, learned mixing coefficients yield 22.82 m versus the static default’s 17.86 m, while greedy soup rejects every addition and ends with ODD-Nagoya alone.The uniform soup is 21.81 m, worse than ODD-Nagoya alone at 17.86 m.Source: D Weight-Averaging Baselines: Learned… · p. 16
  • The uniform soup loses to the best single expert except in two cells, while TIES-Merging is statistically indistinguishable from the uniform soup and loses to the best expert.The fleet-mean inter-task gap is 1.4–5.4 times larger than the interpolation advantage; TIES yields 22.4–22.5 m versus 21.8–22.3 m for the uniform soup, against 17.86 m for the best expert.Source: D Weight-Averaging Baselines: Learned… · pp. 17–18

E Shared-Adapter Repair: Per-Expert Results and Perturbation Ablation

Shared-module repair tests perturbed training scenes, while distillation compresses a soft expert mixture into a regional adapter. Joint replay unfreezes the adapter stack and trades target-region performance against forgetting on training ODDs.

E Shared-Adapter Repair: Per-Expert Results and Perturbation Ablation

  • Disabling SE(2) rotation costs the most; Bezier warp and boundary drop contribute less and by similar amounts.The earlier single-seed claim that boundary drop costs nothing was retracted because its no-boundary-drop Morikoro result had a 13.96 m seed-to-seed standard deviation.Source: E Shared-Adapter Repair: Per-Expert… · pp. 18–19, Table 16 · p. 19

F.1 Regional Adapter via Distillation

  • The distilled student follows coarse route geometry but beats the static default only on Ishiodai, and is worse on Knowledge Hub and Morikoro.The student is a rank-4 LoRA trained on 56 scenes and evaluated on disjoint lanelets.Source: §F.1 · p. 19, Table 17 · p. 19

G.1 The Joint-Replay Frontier

  • With a budget of 20 target-region scenes, joint replay achieves 2.24/2.77/7.78 m on Knowledge Hub/Ishiodai/Morikoro on the same evaluation pool.These are joint-replay results from a single checkpoint.Source: §F.1 · p. 20
  • Joint replay unfreezes the full adapter stack and trains on 200 training-ODD scenes plus a target-region pool at a 10:1 ratio for 500 optimiser steps.The pool-size setting moves the frontier between the frozen adapter stack and full retraining.Source: §G.1 · p. 20
  • With the inherited shared module, 20 target scenes give 15.43 m with −0.04 m training-ODD route-deviation change; training the shared module from scratch reaches 5–7 m but costs +11–16 m regression.These results use the harder rank-sweep evaluation slice with 30 training scenes per map and are not comparable to main-text values.Source: §G.1 · p. 20

H.1 Joint-Replay Sample Complexity

Joint-replay sample-complexity experiments vary the target-region pool while holding the training-ODD pool and optimiser steps fixed. The reported results indicate that larger target pools stabilize outcomes across seeds, while increased model capacity alone does not close the gap to distillation.

H.1 Joint-Replay Sample Complexity

  • The joint-replay sweep uses target-region pools of 5, 10, 20, 40, or 80 scenes, a fixed 200-scene training-ODD pool, and 500 optimiser steps.Evaluation uses the corresponding seed-7 pool of 24 disjoint target-region scenes.Source: §H.1 · p. 21, Table 18 · p. 21
  • At n=20, the three-seed standard deviation is 1.23 m, down 4.2× from n=5, and the authors identify n≥20 as the practical stable budget.At n=5, the single-seed minimum is 6.76 m but the three-seed mean is 9.79 m; at n=40, the mean is 5.67 ± 1.06 m.Source: §H.1 · pp. 21–22

I.1 Capacity under Reward Supervision

  • A rank-128 Skyline LoRA (≈2.5 M trainable parameters, ∼32× the distilled adapter), trained by reward backpropagation for 2000 steps, does not close the gap to the small adapter distilled with MSE loss; the bottleneck is supervision, not capacity.Source: §I.1 · p. 22

J.1 Cross-City Evaluation on a Public Benchmark

The nuPlan leave-one-city-out evaluation tests whether fine-tuning on three cities transfers to a held-out fourth city. Open-loop lateral offsets improve on held-out cities, including distant Singapore, while the transfer matrix shows varying retention across cities.

J.1 Cross-City Evaluation on a Public Benchmark

  • The leave-one-city-out protocol trains on the other three cities and evaluates on a fixed pool of 50 scenes per city.Singapore is the only left-hand-traffic city; the other three cities use right-hand traffic.Source: §J.1 · p. 22
  • Every held-out diagonal entry improves on the base planner, and every training-city entry also improves on the base.The three-seed evaluation uses 50 fixed scenes per city; standard deviations are below 1 m for every entry.Source: §J.1 · p. 22, Table 20 · p. 23
  • Singapore’s held-out lateral offset falls from 8.64 to 5.60 ± 0.78 m (−35%) without Singapore data.The city is 5,000–15,000 km from the three training cities, and the experiment does not test joint replay.Source: §J.1 · p. 23, Table 20 · p. 23
  • Held-out diagonal entries are within 80%–99% of entries for cities seen during training, consistent with near-lossless transfer across North American cities and partial transfer to Singapore.This is the single-seed transfer matrix result.Source: §J.1 · p. 23, Table 21 · p. 23

K Routing Feature Ablations

The routing ablations show that richer features and several scene-level selectors do not reliably identify the best expert, while the appendix reports a separate DPO+L2+SAM adapter study. In that study, the selected configuration has lower comfort-violation rates than the reward-back-prop family on the reported validation comparison.

K Routing Feature Ablations

  • At 38 feature dimensions, validation accuracy reaches 1.000 while oracle routing accuracy falls from 0.375 to 0.250 and fleet mean rises from 31.0 to 36.9 m.The static default’s advantage comes from always using the expert of the nearest training ODD, not per-scene discrimination.Source: K Routing Feature Ablations · pp. 23–24

L Hyperparameters and Training Recipes

M Details of the Routing Schemes

  • The scene-conditioned MLP doubles the static default’s offset on Ishiodai and Morikoro, while self-consistency favors the shared module and denoising-loss correlation is 0.172.The conformal corridor abstains on every scene at T=5 m and adds no triage signal at T=15 m.Source: M Details of the Routing Schemes · p. 26
  • The reward-supervised gate places 0.967 of its weight on ODD-Nagoya, reward-as-scorer routes 20 of 24 scenes to ODD-Shalun, and the oracle reduces mean offset from 16.04 to 15.15 m.On the 24-scene pool, joint replay reduces mean offset from 16.04 to 6.95 m; classifiers reach only majority-class accuracy.Source: M Details of the Routing Schemes · p. 26

N Conformal Calibration Details

  • The pooled exchangeability test rejects equality of training-ODD and target-region calibration residuals, with an observed difference of +18.15 m and p=0.0035.Per-expert tests reject for ODD-Hongo and ODD-Shalun, but not ODD-Kashiwa or ODD-Nagoya.Source: N Conformal Calibration Details · pp. 26–27, Table 25 · p. 26

O.2 Hybrid DPO + L2 Imitation Loss and SAM Outer Loop

  • The separate deployment-oriented study combines DPO with an L2 imitation anchor and an outer SAM loop; accepted preference pairs require a driverlike chosen trajectory and a rejected trajectory violating at least one criterion.Across the sweep, non-zero anchor weights from 0.3 to 0.9 are stable, pure DPO collapses, and comfort varies by up to 30 percentage points across seeds.Source: O Real-Vehicle Runs and a Separate DPO… · p. 27, §O.2 · p. 28

O.1 JAMA-Inspired Preventability Proxy

  • The JAMA-inspired V3 proxy scores trajectories across strict-safety, driverlike, and comfort layers, but is not a full implementation of JAMA Ver. 4.0 §6.It omits the required brake-delay simulator and judge of Responder and Initiator roles; its criteria include author-defined thresholds.Source: §O.1 · p. 27

O.3 Final Comparison (Phase 4) on 500 Validation Scenes

  • Comfort-violation rates are 79–88% for the reward-back-prop family and 0% for DPO+L2+SAM with the same LoRA architecture.The longer Phase 3 run rises to 5.20%, which the authors attribute to overfitting rather than checkpoint noise.Source: §O.3 · p. 28

O.4 Per-Layer Ablation and a Training-Free Fix

The appendix identifies a single adapter layer whose removal sharply reduces comfort violations, then examines safety, comfort, and transfer limits across adversarial tests and vehicle runs. It also reports that the nearest expert is not always the best fixed expert in additional regions.

O.4 Per-Layer Ablation and a Training-Free Fix

  • 7.4% vs 81.4%: zeroing preproj.fc1’s LoRA update at inference lowers Reward-back-prop’s comfort-violation rate on the n=500 validation pool.The layer is the first projection of noisy input xt; its gradient accumulates over 80 sampling steps and pulls weights off the diffusion manifold. The resulting design rule is to keep DiT input and output projections out of the LoRA target set.Source: §O.4 · pp. 28–29

O.5 Collapse of DPO Without the L2 Anchor

  • Without the imitation anchor, pure DPO triggers 100% of every V3 layer in every scenario bucket; any non-zero anchor weight prevents this collapse.At wimit=0, median sampled jerk is 3627 m/s3 and mean predicted speed is 46.6 m/s; with wimit=0.1, comfort violations fall to 53.3%.Source: §O.5 · p. 30

O.6 Greedy Weight Soup over the Sweep Variants

  • The greedy soup over 12 sweep variants reduces to DPO+L2+SAM because no candidate strictly improves on its 0% comfort-violation rate.The metric is saturated at n=500.Source: §O.6 · p. 30

O.8 Scoring Real-Vehicle Recordings with V3

  • The same adapter produced a 47.7% comfort-violation rate versus 37.7% for a human driver in one vehicle run, while a second run crept 72 m and stopped.The authors add a whole-recording progress check to flag runs that are nearly stationary.Source: §O.8 · p. 31

O.9 Comparison with the Production DPO LoRA

  • In simulation, DPO+L2+SAM matches the production DPO LoRA on hard safety at 0% and is slightly better on comfort, 0.00% versus 0.40%.The proposed configuration uses automatically assembled preference pairs without human preference labels; the comparison is limited to simulation.Source: §O.9 · pp. 31–32

O.10 Adversarial Scenarios from AWSIM-Script

  • On cut-in scenarios, Reward-back-prop has a 39.8% collision rate versus 17.0% for the base planner, while DPO+L2+SAM records 18.5%.The TTCp25 heatmap is more dangerous for Reward-back-prop than DPO+L2+SAM, with Reward-LoRA (surgical zero) between them; collision values are relative comparisons, not absolute rates.Source: §O.10 · pp. 33–34

O.11 Training with AWSIM-Script Adversarial Pairs

  • With 500 adversarial pairs, cut-in collisions fall to 17.07% from 18.51%, cut-out collisions fall to 0%, but deceleration remains at 51.12%.The chosen trajectories brake at no more than 4 m/s2, which is too weak for an NPC braking at 1 G; natural-data comfort returns to 0% by epoch 6 and stays there through epoch 8.Source: §O.11 · p. 34

R Regions Where the Nearest Expert Is Not ODD-Nagoya

Loading 2610.31095v1…