PersonaDrive: Controllable Trajectory Prediction with Multi-Dimensional Driving Personas

Chan Lee1, Kimin Yun2, Yuseok Bae2, Seong Tae Kim1,†, Jung Uk Kim1,†
1Kyung Hee University, Yong-in, South Korea
{cksdlakstp12, st.kim, ju.kim}@khu.ac.kr
2ETRI, Daejeon, South Korea
{kimin.yun, baeys}@etri.re.kr
Paper Code PCT Dataset Checkpoint

ECCV 2026

Abstract

Recent trajectory prediction and end-to-end driving methods are robust but offer little control over driving behavior, and existing benchmarks cover at most a single urgency spectrum. We introduce the Persona-Conditioned Trajectory (PCT) dataset, which decomposes driving persona into two axes, Temporal Urgency and Ride Comfort, and pairs each of the resulting nine personas with natural-language descriptions and trajectories. We also propose PersonaDrive, a framework that learns personas from language and generates persona-specific trajectories through a Persona-Conditioned Anchor Transform (PCAT) and Persona-Conditioned Multi-Modal Fusion (PCMF), trained with a Hierarchical Guide Loss and an Axis-Decomposed Diversity Loss. PersonaDrive consistently outperforms baselines across all nine personas, with the largest gains in multi-dimensional scenarios.

Motivation

A single urgency spectrum (emergency / normal / relaxed) cannot tell apart personas that are equally urgent but need different dynamics. A paramedic protecting a critical patient wants a smooth, jerk-free ride, while a firefighter racing to a blaze accepts aggressive maneuvers. We therefore decompose persona into two orthogonal axes, Temporal Urgency (how fast to progress) and Ride Comfort (how smoothly), with three levels each. This yields nine personas that natural language can express beyond a one-hot label.

Comparison of existing planners without persona input, one-dimensional persona planners, and PersonaDrive with a multi-dimensional (Urgency x Comfort) persona grid

Figure 1. Unlike planners with no persona input (a) or a single urgency axis (b), PersonaDrive (c) controls trajectories over a 3×3 grid of Temporal Urgency and Ride Comfort.

Persona-Conditioned Trajectory (PCT) Dataset

Each persona is a natural-language passenger request that encodes urgency and comfort implicitly, covering the full 3×3 grid without explicit axis labels. A four-stage pipeline produces nine trajectory and text pairs for every scene in OpenScene. GPT-4o-mini first generates candidates, rule-based checks validate them, GPT-4o then scores them as an LLM judge, and failed samples are regenerated with the failure reason fed back into the prompt. Humans identify the correct persona with 85.7% accuracy, and independent re-judging with Claude and Gemini agrees strongly (κ = 0.978 and 0.804).

Four-stage PCT dataset generation pipeline: candidate generation, rule-based validation, LLM-as-a-Judge, and failure-aware regeneration

Figure 2. Four-stage dataset generation pipeline: candidate generation, rule-based validation, LLM-as-a-Judge scoring, and failure-aware regeneration.

Radar-chart statistics of the PCT dataset along the urgency axis (longitudinal dynamics) and comfort axis (handling and lateral dynamics)

Figure 3. Dataset statistics. Urgency governs speed and travel distance (a), while comfort governs jerk and yaw rate (b).

Method

PersonaDrive takes camera, LiDAR, and a persona text encoded by a frozen MiniLM. PCAT modulates the anchor bank, using an urgency scalar αu to scale travel distance and a per-timestep comfort vector αc to shape the trajectory. PCMF then fuses BEV, ego, persona, and anchor queries into persona-aware features. Two objectives supervise training on top of the DiffusionDrive baseline: a Hierarchical Guide Loss that enforces axis-aligned orderings (higher urgency means longer trajectories, and lower comfort allows higher jerk) and an Axis-Decomposed Diversity Loss that prevents diagonal mode collapse.

Overall architecture of PersonaDrive: frozen text encoder, PCAT anchor modulation, PCMF fusion, and the trajectory decoder with hierarchical guide and axis-decomposed diversity losses

Figure 4. Overall architecture of PersonaDrive, built on PCAT anchor modulation and PCMF fusion.

Results

On the NAVSIM closed-loop benchmark (all baselines share the same frozen text encoder), PersonaDrive achieves the lowest ADE/FDE and the highest PDMS, winning ADE/FDE in all nine persona cells.

Main results table: PersonaDrive achieves the best average ADE (2.39), FDE (3.71), and PDMS (57.7) against ten baselines on NAVSIM

Table 1. Main results on the NAVSIM navtest split, averaged over the nine personas.

Ablation table: adding PCAT, PCMF, and the axis-decomposed diversity loss progressively improves ADE, FDE, and PDMS over the baseline

Table 2. Ablation study on PCAT, PCMF, and the Axis-Decomposed Diversity Loss (𝐿AD).

Qualitative Results

Higher urgency yields longer trajectories, and higher comfort yields smoother motion. In each figure, (a) shows the nine predicted trajectories over the 3×3 persona grid and (b) shows the natural-language passenger request that conditioned each cell.

Qualitative results scene 1: nine predicted trajectories over the urgency-comfort grid with the conditioning passenger requests

Figure 5. Qualitative results (scene 1). Rows vary Temporal Urgency and columns vary Ride Comfort.

Qualitative results scene 2: nine predicted trajectories over the urgency-comfort grid with the conditioning passenger requests

Figure 6. Qualitative results (scene 2).

Qualitative results scene 3: nine predicted trajectories over the urgency-comfort grid with the conditioning passenger requests

Figure 7. Qualitative results (scene 3).

BibTeX

@misc{lee2026personadrivecontrollabletrajectoryprediction,
      title={PersonaDrive: Controllable Trajectory Prediction with Multi-Dimensional Driving Personas}, 
      author={Chan Lee and Kimin Yun and Yuseok Bae and Seong Tae Kim and Jung Uk Kim},
      year={2026},
      eprint={2608.15230},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.15230}, 
}