cs.AIMar 31, 2026

C-TRAIL: A Commonsense World Framework for Trajectory Planning in Autonomous Driving

Authors: Zhihong Cui, Haoran Tang, Tianyi Li, Yushuai Li, Peiyuan Guan, Amir Taherkordi, Tor Skeie

Organizations: the Department of Informatics, University of Oslo, Norway · the Department of Computing, Hong Kong Polytechnic University, Hong Kong, China · the Department of Computer Science, Aalborg University, Denmark

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Trajectory planning for autonomous driving increasingly leverages large language models (LLMs) for commonsense reasoning, yet LLM outputs are inherently unreliable, posing risks in safety-critical applications. We propose C-TRAIL, a framework built on a Commonsense World that couples LLM-derived commonsense with a trust mechanism to guide trajectory planning. C-TRAIL operates through a closed-loop Recall, Plan, and Update cycle: the Recall module queries an LLM for semantic relations and quantifies their reliability via a dual-trust mechanism; the Plan module injects trust-weighted commonsense into Monte Carlo Tree Search (MCTS) through a Dirichlet trust policy; and the Update module adaptively refines trust scores and policy parameters from environmental feedback. Experiments on four simulated scenarios in Highway-env and two real-world levelXData datasets (highD, rounD) show that C-TRAIL consistently outperforms state-of-the-art baselines, reducing ADE by 40.2%, FDE by 51.7%, and improving SR by 16.9 percentage points on average. The source code is available at https://github.com/ZhihongCui/CTRAIL.

Figures & tables

Fig. 1: The commonsense world paradigm for trajectory planning in dynamic environments
Fig. 2: The C-TRAIL, A Commonsense World Framework for Trajectory Planning in Autonomous Driving
Fig. 4: Examples of correct and incorrect LLM outputs.
CategorySetting
Simulation Environment
SimulatorHighway-env [ 46 ]
ScenariosHighway, Merge, Roundabout,
 Intersection
Action space ∣A∣|\mathcal{A}|{IDLE, FASTER, SLOWER,
 LANE_LEFT, LANE_RIGHT}
TABLE I: Experiment settings and hyperparameters for C-TRAIL.
ScenarioMetricDQNMCTSGRADLMTrajDiLuLangMPCGPT-DrC-TRAIL (3.5)C-TRAIL (4o)p -value
Seen Environments
HighwayADE ↓\downarrow2.87 ±\pm 0.312.63 ±\pm 0.262.51 ±\pm 0.251.32 ±\pm 0.151.15 ±\pm 0.151.25 ±\pm 0.181.20 ±\pm 0.170.77 ±\pm 0.120.68 ±\pm 0.113.1e-5
FDE ↓\downarrow5.12 ±\pm 0.274.82 ±\pm 0.254.61 ±\pm 0.223.24 ±\pm 0.182.81 ±\pm 0.173.05 ±\pm 0.202.95 ±\pm 0.190.92 ±\pm 0.200.80 ±\pm 0.098.7e-6
SR(%) ↑\uparrow25.4 ±\pm 3.329.5 ±\pm 2.531.2 ±\pm 2.068.3 ±\pm 2.473.2 ±\pm 1.970.8 ±\pm 2.271.5 ±\pm 2.186.6 ±\pm 2.588.3 ±\pm 1.65.2e-6
MergeADE ↓\downarrow3.72 ±\pm 0.253.48 ±\pm 0.223.36 ±\pm 0.182.17 ±\pm 0.321.83 ±\pm 0.222.04 ±\pm 0.281.95 ±\pm 0.261.27 ±\pm 0.580.93 ±\pm 0.354.8e-3
FDE ↓\downarrow5.93 ±\pm 0.215.61 ±\pm 0.185.47 ±\pm 0.163.89 ±\pm 0.373.07 ±\pm 0.363.52 ±\pm 0.393.35 ±\pm 0.341.62 ±\pm 0.281.42 ±\pm 0.321.4e-4
TABLE II: Performance Comparison in Seen and Unseen Environments.
VariantsHighwayMergeRoundaboutIntersection
ADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrow
w/o Trust1.10 ±\pm 0.271.45 ±\pm 0.3172.4 ±\pm 5.61.55 ±\pm 0.252.34 ±\pm 0.4270.1 ±\pm 4.42.30 ±\pm 0.323.40 ±\pm 0.5268.5 ±\pm 4.92.85 ±\pm 0.074.75 ±\pm 0.0965.7 ±\pm 3.7
w/o Dirichlet policy1.45 ±\pm 0.301.82 ±\pm 0.3763.5 ±\pm 6.21.85 ±\pm 0.262.96 ±\pm 0.4360.2 ±\pm 4.72.78 ±\pm 0.363.91 ±\pm 0.5658.7 ±\pm 5.43.12 ±\pm 0.115.45 ±\pm 0.1355.3 ±\pm 4.3
w/o Adaptive update0.88 ±\pm 0.251.12 ±\pm 0.2682.3 ±\pm 5.31.20 ±\pm 0.181.85 ±\pm 0.2981.0 ±\pm 3.61.70 ±\pm 0.252.30 ±\pm 0.4479.8 ±\pm 3.52.05 ±\pm 0.053.95 ±\pm 0.0877.6 ±\pm 2.9
C‑TRAIL (GPT‑4o)0.68 ±\pm 0.110.80 ±\pm 0.0988.3 ±\pm 1.60.93 ±\pm 0.351.42 ±\pm 0.3287.1 ±\pm 1.81.38 ±\pm 0.251.75 ±\pm 0.7085.2 ±\pm 2.21.62 ±\pm 0.733.05 ±\pm 0.3484.2 ±\pm 3.2
TABLE III: Ablation Study of Key Modules on Planning Performance.
Fig. 5: Trust score dynamics ( CtC_{t} ) under normal and error injection conditions on highway . Errors are injected at t=4t{=}4 .
  No filteringckin\boldsymbol{c_{\text{kin}}} onlyDual trust (full)
HighwayRPA ↑\uparrow68.40 ±\pm 2.1579.60 ±\pm 1.8589.30 ±\pm 1.20
SR(%) ↑\uparrow74.80 ±\pm 2.3082.70 ±\pm 1.9588.30 ±\pm 1.35
MergeRPA ↑\uparrow65.10 ±\pm 2.4576.30 ±\pm 2.1086.80 ±\pm 1.45
SR(%) ↑\uparrow71.50 ±\pm 2.6080.30 ±\pm 2.2087.10 ±\pm 1.50
RoundaboutRPA ↑\uparrow61.20 ±\pm 2.7074.80 ±\pm 2.2585.70 ±\pm 1.55
SR(%) ↑\uparrow70.30 ±\pm 2.8578.90 ±\pm 2.4085.20 ±\pm 1.70
TABLE IV: LLM Output Quality Before and After Trust Filtering.
ScenarioMethodADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrow
Highway (Normal)GRAD2.51 ±\pm 0.254.61 ±\pm 0.2231.2 ±\pm 2.0
DiLu1.15 ±\pm 0.152.81 ±\pm 0.1773.2 ±\pm 1.9
C-TRAIL0.68 ±\pm 0.110.80 ±\pm 0.0988.3 ±\pm 1.6
Highway (Aggressive)GRAD3.18 ±\pm 0.335.42 ±\pm 0.3022.5 ±\pm 2.8
DiLu1.72 ±\pm 0.223.65 ±\pm 0.2458.4 ±\pm 3.1
C-TRAIL0.92 ±\pm 0.151.18 ±\pm 0.1479.6 ±\pm 2.4
TABLE V: Performance Under Normal vs. Aggressive NPC Behavior.
γdiri\gamma_{\text{diri}}HighwayMergeRoundaboutIntersection
ADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrowADE ↓\downarrowFDE ↓\downarrowSR(%) ↑\uparrow
0.10.82 ±\pm 0.140.98 ±\pm 0.1284.5 ±\pm 2.31.15 ±\pm 0.391.72 ±\pm 0.3783.2 ±\pm 2.41.62 ±\pm 0.282.08 ±\pm 0.7481.3 ±\pm 2.81.95 ±\pm 0.783.52 ±\pm 0.4079.8 ±\pm 3.6
0.30.68 ±\pm 0.110.80 ±\pm 0.0988.3 ±\pm 1.60.93 ±\pm 0.351.42 ±\pm 0.3287.1 ±\pm 1.81.38 ±\pm 0.251.75 ±\pm 0.7085.2 ±\pm 2.21.62 ±\pm 0.733.05 ±\pm 0.3484.2 ±\pm 3.2
0.50.75 ±\pm 0.130.91 ±\pm 0.1186.1 ±\pm 2.01.05 ±\pm 0.371.58 ±\pm 0.3584.8 ±\pm 2.11.51 ±\pm 0.271.93 ±\pm 0.7283.0 ±\pm 2.51.78 ±\pm 0.753.28 ±\pm 0.3781.7 ±\pm 3.4
0.70.91 ±\pm 0.161.08 ±\pm 0.1482.7 ±\pm 2.51.28 ±\pm 0.411.89 ±\pm 0.3980.5 ±\pm 2.81.78 ±\pm 0.312.35 ±\pm 0.7878.4 ±\pm 3.12.15 ±\pm 0.803.85 ±\pm 0.4376.3 ±\pm 3.8
0.91.12 ±\pm 0.191.35 ±\pm 0.1878.2 ±\pm 3.11.52 ±\pm 0.442.24 ±\pm 0.4275.8 ±\pm 3.32.08 ±\pm 0.352.78 ±\pm 0.8273.6 ±\pm 3.72.58 ±\pm 0.854.32 ±\pm 0.4870.5 ±\pm 4.2
TABLE VI: Sensitivity of Trust Calibration Rate γdiri\gamma_{\text{diri}} on Planning Performance.
  Density 2.0Density 3.0
ScenarioLanesADE ↓\downarrowFDE ↓\downarrowADE ↓\downarrowFDE ↓\downarrow
Highway40.68 ±\pm 0.110.80 ±\pm 0.090.74 ±\pm 0.120.88 ±\pm 0.11
50.71 ±\pm 0.120.84 ±\pm 0.110.78 ±\pm 0.130.92 ±\pm 0.12
Merge40.93 ±\pm 0.351.42 ±\pm 0.321.05 ±\pm 0.421.57 ±\pm 0.37
50.97 ±\pm 0.361.48 ±\pm 0.331.11 ±\pm 0.381.64 ±\pm 0.35
Roundabout21.38 ±\pm 0.251.75 ±\pm 0.701.45 ±\pm 0.281.87 ±\pm 0.75
TABLE VII: Robustness to Lane Count and Vehicle Density.
   HighwayRoundab.highDrounD
LMTrajInfer (s) 1.56 ±\pm 0.141.73 ±\pm 0.201.61 ±\pm 0.161.78 ±\pm 0.22
Mem (MB) 276 ±\pm 4295 ±\pm 4278 ±\pm 5298 ±\pm 5
DiLuInfer (s) 15.39 ±\pm 0.5017.81 ±\pm 0.3715.72 ±\pm 0.4818.05 ±\pm 0.42
Mem (MB) 613 ±\pm 19662 ±\pm 20621 ±\pm 20670 ±\pm 21
LangMPCInfer (s) 12.35 ±\pm 0.3913.42 ±\pm 0.3612.52 ±\pm 0.3813.58 ±\pm 0.40
Mem (MB) 485 ±\pm 14510 ±\pm 15490 ±\pm 14515 ±\pm 15
TABLE VIII: Computational Efficiency of LLM-Based Methods.
Fig. 6: Successfully executed planned trajectories across four scenarios. The green box denotes the ego vehicle, which follows the planned 5-step trajectory. Blue boxes represent surrounding vehicles controlled by the simulator.
Fig. 7: Failed planned trajectories resulting in collisions, illustrating three failure types: policy error (a), model error (c), and mapping error (d).
ScenarioPolicy ErrorModel ErrorMapping Error
Highway52.30 ±\pm 3.1228.60 ±\pm 2.7119.10 ±\pm 2.38
highD50.80 ±\pm 3.4230.40 ±\pm 2.9318.80 ±\pm 2.51
Roundabout45.10 ±\pm 3.2435.80 ±\pm 3.0519.10 ±\pm 2.33
rounD42.60 ±\pm 3.4838.50 ±\pm 3.1518.90 ±\pm 2.58
Average47.70 ±\pm 3.3233.33 ±\pm 2.9619.00 ±\pm 2.45
TABLE IX: Failure Breakdown (%) in Seen Environments.
MethodRecallPlanUpdate
DiLu [ 47 ]✓\checkmarkΔ\DeltaΔ\Delta
LC-LLM [ 61 ]✓\checkmark✓\checkmark×\times
Voyager [ 36 ]✓\checkmarkΔ\Delta✓\checkmark
TrafficSafetyGPT [ 22 ]✓\checkmark×\times×\times
Wen et al. [ 62 ]×\times✓\checkmark×\times
MBAPPE [ 63 ]×\times✓\checkmark×\times
TABLE X: Comparison of representative methods across the Recall, Plan, and Update stages in trajectory planning. ✓\checkmark : the method explicitly designs a dedicated module or mechanism for the stage; Δ\Delta : the stage is partially addressed as a byproduct of the main design, without a dedicated or complete mechanism; ×\times : the stage is not addressed.
Figure 17
Figure 18
Figure 19
Figure 20
Figure 21
Figure 22
Figure 23

Explore similar work

CardsList

  1. Reason--Imagine--Act: Closed-Loop LLM Decision Making with World Models for Autonomous Driving

    May 19, 2026Zhengqi Sun, Yiwen Sun, Boxuan Liu +3Language Model-Based ControlWorld Models

    Department of Information Management, Peking University, Beijing 100871, China · School of Intelligence Science and Technology, Peking University, Beijing 100871, China · State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing 100080, China +3

  2. LWDrive: Layer-Wise World-Model-Guided Vision-Language Model Planning for Autonomous Driving

    Jun 29, 2026Chen Yang, Yuhao Wei, Ze Xu +6VLMs for Autonomous DrivingEnd-to-End Autonomous Driving

    Chongqing University

  3. GAPL: Grounded Action-effect Policy Learning for LLM-Based Trajectory Planning

    Aug 18, 2026Zhihong Cui, Hengyu Liu, Zhangkai Wu +5Autonomous Driving PlanningLarge Language Model-Based Robot Planning

    Department of Informatics, University of Oslo, Oslo, Norway · Department of Computer Science, Aalborg University, Aalborg, Denmark · University of Sydney, Sydney, Australia +2