SeekLifeShallDie/PCT-Dataset

Persona-Conditioned Trajectory (PCT) Dataset

The PCT Dataset is the benchmark introduced in PersonaDrive: Controllable Trajectory Prediction with Multi-Dimensional Driving Personas (ECCV 2026).

Unlike prior benchmarks that provide either no persona-conditioned annotations or only a single urgency spectrum (emergency / normal / relaxed), PCT decomposes driving personas along two independent axes:

• Temporal Urgency — how quickly the driver wants to reach the goal
• Ride Comfort — how smooth / aggressive the motion is

Three levels of each axis combine into a 3 × 3 grid of nine personas. Every persona is paired with a natural-language driver intent and a corresponding future trajectory, enabling models to learn driving personas from language and generate persona-specific trajectories.

The PCT annotations are built on top of the NAVSIM / OpenScene (nuPlan) scenes. This dataset provides only the persona-conditioned labels; you need the original NAVSIM sensor/log data to use them for training and evaluation.

The 9 Personas

Personas are keyed as U{x}_C{y} where x, y ∈ {H, M, L}:

	Comfort H	Comfort M	Comfort L
Urgency H	UH_CH	UH_CM	UH_CL
Urgency M	UM_CH	UM_CM	UM_CL
Urgency L	UL_CH	UL_CM	UL_CL

Dataset Structure

The archive PersonaDrive_dataset.tar.gz extracts to a single directory:

per_scene/
├── 847c572c52cd53b0.json
├── 801b1e1314c55e0a.json
└── ...                       # 115,434 scene files

Each file is named by its NAVSIM scene token and contains the nine personas plus a _selection field:

{
  "UH_CL": {
    "trajectory": [[0.0, 0.0, 0.4382], [1.0339, 0.4845, 0.1683], ...],
    "params": { "maneuver": "follow", "speed_factor": 1.95, "lane_offset": 1.2 },
    "user_intent": "Please, can we go faster? There's a flood at my house ..."
  },
  "UH_CM": { ... },
  ...
  "UL_CL": { ... },
  "_selection": { ... }
}

Field	Description
trajectory	Future waypoints in the ego frame, each [x, y, heading] (meters, radians).
params.maneuver	High-level maneuver (e.g. follow, lane_change, ...).
params.speed_factor	Speed scaling relative to the nominal trajectory (urgency axis).
params.lane_offset	Lateral offset in meters (comfort / style axis).
user_intent	Natural-language description of the driver's intent for this persona.

Usage

import json, tarfile

# Extract
# tar xzf PersonaDrive_dataset.tar.gz

with open("per_scene/847c572c52cd53b0.json") as f:
    scene = json.load(f)

persona = scene["UH_CL"]
print(persona["user_intent"])
print(persona["trajectory"])   # list of [x, y, heading]

For full training and evaluation, follow the instructions in the PersonaDrive repository — environment setup, dataset preparation, and the training/eval scripts.

Results

PersonaDrive trained on PCT, averaged over all nine personas:

Method	Avg. ADE ↓	Avg. FDE ↓	Avg. PDMS ↑
LTF (TPAMI'22)	2.49	3.87	52.0
Transfuser (TPAMI'22)	2.53	3.96	51.9
UniAD (CVPR'23)	2.47	3.83	52.5
DiffusionDrive (CVPR'25)	3.93	5.79	55.3
PersonaDrive (Ours)	2.39	3.71	57.7

License

Released under CC BY-NC 4.0 (non-commercial research use). The PCT annotations derive from NAVSIM / OpenScene (nuPlan); please also comply with the original nuPlan and NAVSIM licenses.

Acknowledgement

Our codes benefits from the excellent NAVSIM, Transfuser, Diffusion Policy, MapTR, VAD, SparseDrive, DiffusionDrive.