Organizations: Shuo Pei, Saihui Zha, Huachun Tan are with the Beijing Institute of Technology, Guangdong, China. · Shuo Pei, Bosun Liang, Chen Sun, and Yong Wang (Corresponding author) are with the Department of Data and Systems Engineering, The University of Hong Kong, Hong Kong, China.
Generative planners are often evaluated by the trajectories they can generate, although closed-loop driving ultimately depends on the single candidate committed to control. We identify this mismatch as the sample-to-execution gap and show that improving proposal quality alone can worsen the executed decision. Here, we introduce ExecDiffuser, a recurrent proposal-to-execution framework that couples candidate generation, trajectory selection, and subsequent replanning through shared execution semantics. Intent-Guided Diffusion Proposal (IGD) assigns hypothesis responsibility and shapes trajectories with execution-conditioned supervision. Execution-Grounded Selection (EGS) learns a candidate-conditioned publication utility, commits one trajectory to control, and returns it as context for the next planning cycle. Across nuPlan, InterPlan, and NAVSIM, ExecDiffuser achieves leading closed-loop performance against strong state-of-the-art baselines. Under a unified five-seed protocol, it obtains an average score of 85.62, improving over Diffusion Planner and FlowDrive by 4.02 and 1.31 points, respectively. On NAVSIM, it achieves 92.6 PDMS under the official evaluation pipeline. Matched counterfactual analysis further shows that ExecDiffuser more reliably converts strong candidates into committed decisions, while prefix interventions confirm that this advantage persists after physical execution and subsequent replanning.
Figures & tables
Fig. 1: A strong candidate can already be available yet remain uncommitted to control. Strengthening proposals alone may raise the available-candidate ceiling while widening candidate-publication regret. ExecDiffuser connects proposal learning and trajectory publication through shared execution semantics: IGD improves candidate availability, while recurrent EGS commits one trajectory to control and returns it to the next planning cycle.
Fig. 2: ExecDiffuser couples proposal learning and publication through a shared online execution interface. IGD assigns and shapes latent hypotheses; EGS filters invalid candidates, ranks the remaining candidates, publishes one trajectory, and returns it as recurrent context.
Method
Val14 NR
Val14 R
Hard NR
Hard R
Test14 NR
Test14 R
Avg.
InterPlan
Results reported in the original papers for context
PDM-Closed [ 20 ]
92.84
92.12
65.08
75.19
90.05
91.63
84.49
–
PlannerRFT [ 9 ]
89.96
84.46
77.16
72.21
–
–
–
–
LAP [ 8 ]
89.37
82.23
78.52
70.53
90.64
85.32
82.77
–
G2DP [ 14 ]
92.26
83.91
80.05
73.91
91.58
85.26
84.50
–
G2DP, no history [ 14 ]
91.08
86.67
78.77
77.61
88.54
88.42
85.18
–
TABLE I: Closed-loop results on nuPlan and InterPlan. NR/R denote non-reactive/reactive simulation. Bold compares methods reproduced under the same five-seed protocol; the remaining rows provide reported-paper context.
Proposal training
Val14 NR/R ↑
Oracle-best ↑
Feasible ratio (%) ↑
All-invalid (%) ↓
IGD + imitation WTA
90.84/83.95
91.36
86.60
2.17
+ resp. + shaping
92.05/86.41
92.63
92.10
0.82
+ specialization
92.37/87.11
92.93
93.40
0.68
TABLE II: Stage-controlled Val14 ablations. Panel (a) fixes the full EGS selector and changes proposal learning; panel (b) freezes the same full proposal generator and changes publication. Rule cost minimizes Eq. ( 5 ), and oracle quantities use the complete candidate set.
Fig. 3: Proposal-only contradiction on Val14. Both proposal generators use the same imitation publication rule. Stronger proposals raise the available-best ceiling, but published quality fails to follow. Values are five-seed means.
Fig. 4: Publication cascade with the proposal generator fixed. Candidate-conditioned scoring, online evidence, hard validity, and recurrent context progressively improve oracle top-1 recall and reduce candidate-publication regret. Values are five-seed means.
Prefix horizon
Unique scenarios / eligible states
Δ score vs. imitation (paired 95% CI)
Δ score vs. rule cost (paired 95% CI)
Recovery rate (%) imitation / EGS
0.5 s
1,118 / 111,820
+0.51[0.33,0.69]
+0.37[0.21,0.53]
84.7 / 87.8
1.0 s
1,118 / 111,406
+0.91[0.66,1.15]
+0.62[0.39,0.84]
82.9 / 88.3
2.0 s
1,117 / 110,736
+1.42[1.08,1.75]
+0.95[0.66,1.23]
79.8 / 87.1
4.0 s
1,115 / 109,284
+1.90[1.44,2.34]
+1.25[0.87,1.62]
75.4 / 85.9
TABLE III: Resumed replanning after prefix intervention. Score differences compare EGS with imitation and rule-cost prefixes; intervals are paired over scenario clusters. Recovery uses the same degraded-state denominator within each horizon.
Fig. 5: Matched-prefix intervention on nuPlan. Different initial publications create different physical states before the same frozen planner resumes.
Method
NC
DAC
TTC
EP
PDMS
DiffusionDrive [ 6 ]
98.2
96.2
94.7
82.2
88.1
CdDrive [ 15 ]
98.7
97.5
95.4
82.7
89.2
ReCogDrive [ 27 ]
97.9
97.3
94.9
87.3
90.8
DiffusionDriveV2 [ 10 ]
98.3
97.9
94.8
87.5
91.2
ExecDiffuser
99.1
98.2
97.4
87.6
92.6
TABLE IV: NAVSIM v1.1 navtest results (%). Baselines are reported-paper values; ExecDiffuser is the five-seed mean.
Fig. 6: Qualitative NAVSIM examples of an open-space lane change, lane-aligned motion after a turn, and navigation around queued traffic. Orange denotes the logged expert reference and cyan the trajectory selected by ExecDiffuser.
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
Configuration
Value
Training data
nuPlan v1.1 training split
Calibration split
Scenario-disjoint 10% of training data
Observation / planning horizon
2.0 s / 8.0 s
Planning frequency
10 Hz
Intent queries / budgets
8 / {2,4,8}
Training / inference diffusion steps
1000 / 25
Appendix
TABLE A1: Training, timing, and reporting configuration.
Distance/area outside the commanded route corridor
More route deviation ↑ cost
elane
Lane-boundary overlap and lane-topology inconsistency
More violation ↑ cost
erisk
Predicted overlap plus clearance and TTC shortfalls
More overlap/shortfall ↑ cost
eprogress
Shortfall from route-aligned progress target
Less progress ↑ cost
ekin
Speed, acceleration, curvature, and yaw-rate exceedance
More exceedance ↑ cost
Appendix
TABLE A2: Orientation of the online execution evidence. These normalized quantities enter the training preference, rule-cost control, and learned evidence encoder with the same penalty direction. Coordinate shaping uses aligned differentiable surrogates defined in Eq. ( A21 ).
Group
Parameter
Fixed value
Hard validity
Look-ahead / outside-footprint ratio
3.0 s / 5%
Drivable validity
Ego-center tolerance
0.30 m
Collision validity
Swept overlap / ego inflation
0.05 m 2 / 0.15 m per side
TTC validity
Threshold / minimum closing speed
1.0 s / 0.5 m/s
Uncertainty inflation
Radius
min(0.80,0.15+0.50σt) m
Training-preference weights
drive/route/lane/risk/prog/kin/comfort
1.00/0.80/0.40/1.50/0.30/0.25/0.20
Appendix
TABLE A3: Fixed online-validity, training, allocation, gate, and MRM parameters.
Method
Params (M)
Added params
FLOPs/cycle (G)
Train h/seed
GPU-h/seed
Five-seed GPU-h
Latency (ms)
Diffusion Planner
43.7
–
111.8
42.6
340.8
1704.0
50.7
FlowDrive
46.2
–
96.4
39.1
312.8
1564.0
47.9
ExecDiffuser
48.9
+5.2
138.6
49.8
398.4
1992.0
59.5
Appendix
TABLE A4: Model size and computation under the same NVIDIA GeForce RTX 4090D environment. FLOPs are measured for batch size one at each method’s default operating point; training cost is reported per seed and over all five seeds.
Seed
Val14 NR
Val14 R
Test14-hard NR
Test14-hard R
Test14 NR
Test14 R
InterPlan
0
92.25
86.82
79.31
74.30
92.32
88.08
38.68
1
92.54
87.32
78.91
74.83
92.05
88.45
37.93
2
92.34
87.17
79.18
74.57
92.21
88.41
38.46
3
92.44
87.00
79.04
74.90
92.02
88.27
38.62
4
92.28
87.24
79.21
74.45
92.20
88.39
38.06
Mean
92.37
87.11
79.13
74.61
92.16
88.32
38.35
Appendix
TABLE A5: Per-seed closed-loop results. The mean row exactly matches the main paper’s closed-loop results table; Std. is the sample standard deviation across five independently trained models.
Fig. A1: Seed-level stability across the seven reported settings. Open circles denote independently trained models, horizontal lines denote arithmetic means, and shaded bands denote one sample standard deviation across the five seeds. Exact values appear in Table A5 .
Comparison
Difference
Paired 95% CI
ExecDiffuser − Diffusion Planner, Val14 NR
+2.50
[1.76, 3.21]
ExecDiffuser − Diffusion Planner, Val14 R
+4.73
[3.68, 5.77]
ExecDiffuser − FlowDrive, Val14 NR
+1.16
[0.44, 1.86]
ExecDiffuser − FlowDrive, Val14 R
+1.74
[0.73, 2.71]
ExecDiffuser − Diffusion Planner, Hard R
+5.39
[4.28, 6.42]
ExecDiffuser − FlowDrive, Hard R
+1.52
[0.57, 2.46]
Appendix
TABLE A6: Scenario-level paired bootstrap with 10,000 resamples and identical scenario seeds.
Oracle-best
Published score
Regret
Seed
Base
IGD
Diff.
Base
IGD
Diff.
Base
IGD
Diff.
0
90.28
91.29
+1.01
89.77
89.56
−0.21
0.51
1.73
+1.22
1
90.41
91.44
+1.03
89.90
89.72
−0.18
0.51
1.72
+1.21
2
90.24
91.33
+1.09
89.75
89.59
−0.16
0.49
1.74
+1.25
3
90.35
91.39
+1.04
89.84
89.66
−0.18
0.51
1.73
+1.22
4
90.37
91.35
+0.98
89.84
89.67
−0.17
0.53
1.68
+1.15
Appendix
TABLE A7: Five-seed consistency of the proposal-only contradiction. Both variants use the same imitation selector. Diff. is proposal-only IGD minus Base. The available-best ceiling improves for every seed, while published quality slightly decreases and regret increases for every seed.
Comparison
Mean diff.
Paired 95% CI
Oracle-best
+1.03
[0.76,1.29]
Published score
−0.18
[−0.34,−0.03]
Regret
+1.21
[0.93,1.48]
Oracle top-1 recall
−6.8
[−8.9,−4.6]
Appendix
TABLE A8: Paired scenario-level bootstrap intervals for proposal-only IGD minus Base. The bootstrap uses 10,000 resamples and preserves method and seed pairing. Recall differences are percentage points.
Seed
Published Scf
Oracle-best
Δcf
Top-1 recall
0
92.20
92.86
0.66
85.9
1
92.46
93.03
0.57
87.2
2
92.28
92.92
0.64
86.4
3
92.37
92.98
0.61
86.9
4
92.24
92.86
0.62
86.6
Mean
92.31
92.93
0.62
86.6
Appendix
TABLE A9: Per-seed counterfactual candidate diagnostics over Ωsel . Oracle-best is taken over all generated candidates, including candidates rejected by online hard validity. Published and oracle-best scores use identical candidate rollouts; all-invalid/MRM states are reported separately. Std. is computed across seeds.
Fig. A2: Stage-controlled proposal learning with the full EGS selector fixed. Execution-conditioned responsibility, differentiable geometry shaping, and specialization progressively improve candidate availability and reactive closed-loop performance. Values are five-seed means.
Fig. A3: Execution-grounded publication cascade with the full proposal generator fixed. Learned scoring, online evidence, hard validity, and adaptive recurrent execution progressively improve reactive and TTC scores while reducing candidate-publication regret. Values are five-seed means.
Temporal strategy
Val14 R
TTC
Plan dev. ↓
Switch ↓
Hazard delay ↓
Hazard success ↑
Avg. speed
Unnec. stop ↓
Route progress ↑
No temporal feedback
86.14
92.89
0.313
13.6
0.16
91.9
6.42
3.42
89.82
Fixed temporal feedback
86.49
93.33
0.229
8.8
0.31
90.8
6.38
3.37
89.91
Adaptive Continuity Gate
87.11
94.06
0.205
7.3
0.18
93.4
6.39
3.19
90.05
Appendix
TABLE A11: Recurrent stability, hazard responsiveness, and execution efficiency on Val14. Hazard delay is in seconds; rate columns are percentages.
Commit horizon
Selector
Published score ↑
Oracle-best ↑
Regret ↓
Top-1 recall (%) ↑
0.5 s
Imitation selector
96.84
97.12
0.28
67.3
0.5 s
Rule-cost selector
96.93
97.12
0.19
74.8
0.5 s
Full EGS selector
97.03
97.12
0.09
88.1
1.0 s
Imitation selector
95.38
95.94
0.56
62.5
1.0 s
Rule-cost selector
95.56
95.94
0.38
72.1
1.0 s
Full EGS selector
95.77
95.94
0.17
87.4
Appendix
TABLE A12: Commit-horizon sensitivity under a fixed candidate set. All selectors operate on the same candidates, so oracle-best is identical across selectors within each horizon. Scores should not be compared directly across horizons because the rollout duration changes. Values follow the same five-seed protocol as the primary evaluation.
Prefix
Initial source
Scenarios
Eligible states
Score ↑
TTC score ↑
Collision (%) ↓
Progress score ↑
0.5 s
Imitation prefix
1,118
111,820
86.44
92.81
0.31
89.61
0.5 s
Rule-cost prefix
1,118
111,820
86.58
93.02
0.27
89.72
0.5 s
EGS prefix
1,118
111,820
86.95
93.41
0.19
89.91
0.5 s
Oracle prefix
1,118
111,820
87.12
93.56
0.17
90.02
1.0 s
Imitation prefix
1,118
111,406
86.05
92.38
0.39
89.34
1.0 s
Rule-cost prefix
1,118
111,406
86.34
92.76
0.32
89.55
Appendix
TABLE A13: Closed-loop branch evaluation after prefix execution and resumed replanning. Each branch starts from the same cloned state, executes the indicated prefix, and then uses the same frozen full ExecDiffuser planner for the remainder of the fixed 8.0-s window. Counts are unique scenario tokens and matched eligible diagnostic states. Bold marks the strongest online prefix source; the retrospective oracle is shown only as a diagnostic upper bound.
Prefix
Common degraded
Imitation
Rule cost
EGS
Oracle
0.5 s
18,734
84.7
85.5
87.8
88.4
1.0 s
21,968
82.9
84.1
88.3
89.0
2.0 s
27,615
79.8
81.6
87.1
88.7
4.0 s
35,902
75.4
78.3
85.9
87.6
Appendix
TABLE A14: Recovery rates (%) after resumed replanning. Within each horizon, all prefix sources are evaluated over the same common degraded-state denominator. Bold marks the strongest online prefix source; the oracle remains retrospective.
Seed
Prefix
Imitation
Rule cost
EGS
Oracle
0
0.5 s
86.34
86.51
86.86
87.04
1
0.5 s
86.55
86.71
87.08
87.22
2
0.5 s
86.39
86.56
86.88
87.06
3
0.5 s
86.47
86.64
86.99
87.15
4
0.5 s
86.45
86.49
86.95
87.12
0
1.0 s
85.97
86.27
86.89
87.14
Appendix
TABLE A15: Seed-level resumed scores for each prefix horizon and source. Each row is an independently trained model evaluated under the matched branch protocol.
Prefix
Comparison
Mean diff.
Paired 95% CI
0.5 s
EGS − imitation
+0.51
[+0.33,+0.69]
EGS − rule cost
+0.37
[+0.21,+0.53]
Oracle − EGS
+0.17
[+0.06,+0.29]
1.0 s
EGS − imitation
+0.91
[+0.66,+1.15]
EGS − rule cost
+0.62
[+0.39,+0.84]
Oracle − EGS
+0.24
[+0.10,+0.39]
Appendix
TABLE A16: Scenario-clustered paired bootstrap intervals for resumed score. Scenario tokens are resampled as clusters while retaining their associated method, seed, and diagnostic-state outcomes.
Fig. A4: Scenario-wise resumed-replanning gains. Each cell reports the five-seed EGS-minus-imitation score difference. Columns show 1.0-, 2.0-, and 4.0-s forced-prefix horizons; values should be compared within a column because longer forced prefixes leave less time for the common planner to recover.
Fig. A5: Stability, fallback coverage, and forecast-dependent failure. (a) The Adaptive Continuity Gate reduces plan switching without the hazard-response delay introduced by fixed temporal feedback; marker area encodes hazard success. (b) Learned forecasting with uncertainty increases coverage of counterfactually required MRM events. (c) False-valid, false-reject, and severe low-TTC rates among false-valid decisions increase in high-risk subsets, especially occluded intersections. Rates and scores are pooled or averaged according to the units defined in Appendix Sec. B .
Fig. A6: Scenario dependence, computation, and parameter boundary. (a) Scenario-wise gains over FlowDrive are largest in unprotected turns, pedestrian or cut-in events, dense traffic, and merge or yield interactions. (b) Learned candidate allocation improves the Val14 reactive-score versus latency frontier relative to fixed budgets and heuristic adaptation. (c) Increasing the uncertainty-aware risk margin improves TTC, but excessive margins reduce route progress and overall reactive performance; 0.50 is the fixed default. Values are five-seed means.
Forecast
minADE
minFDE
Val14 R
TTC
False-valid
False-reject
Constant velocity
1.10
2.36
86.02
92.76
4.80
3.20
Learned deterministic
0.82
1.72
86.66
93.51
3.50
3.40
Learned + uncertainty
0.82
1.72
87.11
94.06
2.70
4.00
Appendix
TABLE A17: Forecast accuracy is measured in meters over the 8.0-s horizon on interaction-relevant agents. Validity rates are percentages.
Fig. A7: Forecast and validity diagnostics. (a) Learned forecasting reduces minADE and minFDE relative to constant velocity. (b) Uncertainty reduces false-valid decisions at the cost of a moderate increase in false rejection. (c) The same uncertainty-aware interface raises required-MRM coverage and reduces missed-MRM cycles. Exact values appear in Tables A17 and A21 .
Budget policy
Candidate order
Avg. K
Latency
Budget regret
Oracle-best
Feasible ratio
Val14 NR
Val14 R
Random matched
Random
5.5
59.4
0.43
92.38
90.81
91.76
86.19
Random matched
Scene-conditioned priority
5.5
59.1
0.27
92.61
91.75
92.04
86.48
Learned budget
Fixed query order
5.5
59.6
0.22
92.69
92.26
92.11
86.77
Learned budget
Random order
5.5
59.3
0.30
92.55
91.94
91.97
86.59
Joint learned allocation
Scene-conditioned priority
5.5
59.5
0.12
92.93
93.40
92.37
87.11
Appendix
TABLE A18: Matched-compute controls disentangle budget prediction, candidate priority, and nested ordering. Every row has Kˉ=5.5 and uses the same 25-step decoder, hardware, and five seeds.
Fig. A8: Normalized latent-intent diagnostics. Used slots are divided by eight; percentage metrics are divided by 100; “no collapse” is one minus the collapse rate. The heatmap complements the exact unnormalized values in Table A19 .
TABLE A21: Forecast-failure-aware MRM analysis. The first, second, fifth, and sixth columns except coverage are percentages of all planning cycles; the final two columns are percentages within missed-MRM events. Coverage is the fraction of counterfactually required MRM events detected online.
Subset
Events
False-valid
False-reject
Collision among false-valid
TTC <1 s among false-valid
All diagnostic states
112,840
2.70
4.00
0.8
4.6
Counterfactual TTC <3 s
6,387
3.91
4.54
1.7
8.9
Counterfactual TTC <2 s
2,264
4.76
5.11
2.5
13.8
Cut-in
1,193
4.24
4.69
2.0
10.7
Pedestrian conflict
742
4.58
5.29
2.3
12.0
Merge/yield
1,527
3.67
4.38
1.2
8.5
Appendix
TABLE A22: Validity errors stratified by counterfactual risk and scenario type, pooled across five seeds. Errors rise in the hardest subsets, motivating uncertainty inflation and explicit reporting rather than a global average alone.
Fig. A9: Non-conservative execution diagnostics on Val14. (a) TTC improves without reducing average speed; marker size encodes fewer unnecessary stops. (b) The full model reduces stopping, unnecessary stopping, high-jerk events, and emergency braking across planning cycles. Values are five-seed means.
Seed
NC ↑
DAC ↑
TTC ↑
Comf. ↑
EP ↑
PDMS ↑
0
99.1
98.3
97.5
100.0
87.6
92.7
1
99.0
97.9
97.1
100.0
87.4
92.2
2
98.9
98.3
97.2
100.0
87.3
92.4
3
99.1
98.3
97.5
100.0
87.7
92.8
4
99.2
98.2
97.7
100.0
87.8
92.9
Mean
99.1
98.2
97.4
100.0
87.6
92.6
Appendix
TABLE A23: Per-run results on the NAVSIM v1.1 navtest split. NC, DAC, TTC, Comf., EP, and PDMS denote no at-fault collision, drivable-area compliance, time-to-collision, comfort, ego progress, and PDM Score, respectively. All values are percentages.
Fig. A10: Seed-level stability across the six NAVSIM metrics. Open circles denote independently trained models, horizontal lines denote arithmetic means, and shaded bands denote one sample standard deviation across five runs. Exact values are reported in Table A23 .
Fig. A11: Five-frame BEV excerpts from three NAVSIM scenes. Time progresses from left to right at 0.5-s intervals, covering 2.0 s from the first to the last displayed frame. (a) Open-space selection in a left-lane-change context. (b) Lane-aligned motion after a left turn. (c) A rightward maneuver around queued traffic. Orange and cyan denote the logged expert and ExecDiffuser trajectories, respectively. The sequences illustrate replanning consistency across consecutive logged states rather than closed-loop execution of the predicted trajectories.
Diffusion probabilistic models can capture the multi-modal, interaction-rich distribution of joint future trajectories in driving scenes. We show that a single pretrained diffusion traffic model can serve two complementary roles in the autonomous driving development loop: as an ego motion planner, and as a controllable generator of safety-critical scenarios for stress-testing the planners. On the planning side, we introduce a Single-Stream Dual-Stream (SSDS) diffusion-transformer decoder that fuses scene context via joint attention rather than late cross-attention, improving closed-loop performance on nuPlan. We further propose Decoupled Annealing Posterior Sampling with Energy (DAPSE), a training-free guidance scheme that injects arbitrary energy functions at the clean-sample level, avoiding the first-order approximation errors while requiring no auxiliary networks. Beyond planning, we leverage the same diffusion model as a controllable scenario generator to create realistic long-tail driving interactions for closed-loop evaluation. Through inference-time guidance, selected agents are steered toward safety-critical behaviors, including aggressive cut-ins, lead-vehicle braking, and combined longitudinal-lateral interactions, while preserving realistic traffic behaviors. Evaluated in closed-loop nuPlan simulations with independent black-box planners, the generated scenarios expose failure modes that remain hidden under standard benchmarks. Although the SSDS-based planner achieves stronger nominal performance, it experiences larger degradation under these challenging scenarios, demonstrating that benchmark superiority does not necessarily translate to robustness. These results demonstrate that a single learned traffic prior can simultaneously improve motion planning and provide a realistic framework for systematic planner robustness evaluation.
Arka Pal, Rajesh Kumar, Hannes Eriksson +4
KTH Royal Institute of Technology, Sweden · Zenseact AB, Sweden · Chalmers University of Technology, Sweden
Autonomous driving planning requires selecting trajectories that remain safe, rule-compliant, and comfortable under complex multi-agent interactions, yet robust closed-loop decision making in planning-critical scenes remains challenging. Existing diffusion- and RL-based planners often rely on generic rewards and use relational structure mainly for representation, limiting how planning-critical rollouts shape policy improvement. We present GRID-Planner, a graph--reward integrated diffusion planning framework that aligns generative planning with structured relational signals. Specifically, GRID-Planner combines graph-aware importance weighting, a time-indexed relational reward surrogate, and an advantage-weighted diffusion objective with on-policy self-distillation, forming a rollout-return-advantage-update loop without explicit likelihoods. Experiments on nuPlan and interPlan show that, compared with the diffusion-only baseline, GRID-Planner reduces collision rate, lane deviation, and jerk by 35.8%, 60.8%, and 27.4%, respectively, and improves the planning score by 3.4%. Project website: https://github.com/yangzhaojason/GRID-Planner.
Zhao Yang, Zezhong Qian, Yinan Shi +2
Xi’an Jiaotong University, Xi’an, China · Cytoderm Intelligent Technology, Beijing, China
World-model planners evaluate candidate trajectories by predicting their future consequences, implicitly relying on the plan-to-motion relationship learned during training. At deployment, a different downstream controller may execute the same planned trajectory differently, creating a planning-execution gap between the frozen model's imagined future and actual execution. A natural remedy is to replace the planned trajectory with its predicted or observed execution. However, we find that better execution fidelity does not necessarily yield better decisions: directly substituting predicted execution trajectories yields only a 0.048 PDMS improvement across 10 unseen controller configurations, far smaller than the gain from our proposed adaptation. This suggests a semantic mismatch between a planned trajectory, which serves as a reference for downstream control, and an executed trajectory, which already describes the resulting post-controller motion. We introduce ExecutionBridge, which infers controller-specific execution characteristics from planned-executed trajectory pairs and uses them to adapt the native ego latent rollout while keeping the original planned trajectory unchanged. The planner, world model, and reward head remain frozen. On NAVSIM, ExecutionBridge consistently improves planning across unseen controller configurations and multiple frozen world-model planners, with substantially larger gains in challenging driving scenes where execution mismatch more strongly affects candidate evaluation.
Danqi Zhao, Keyu Chen, Wenchao Sun +6
School of Vehicle and Mobility, Tsinghua University, Beijing, China. · Shanghai Technical Center, DENSO (China) Investment Co., Ltd., Shanghai, China. · DENSO (China) Investment Co., Ltd., China.