JojoZhu/cosmos-trajectory-synthesis

COSMOS Synthetic Traffic Trajectory Dataset

Version: 2026.05-v1 Paper: Language-Guided Traffic Simulation and Rare Event Synthesis for an Urban Intersection (CVPRW 2026) Previous version (2024 workshop): https://huggingface.co/datasets/harrypotter1501/cosmos-trajectory

Overview

Synthetic multi-agent traffic trajectories at a fixed urban intersection, generated by the COSMOS pipeline. Each scene is created from a natural language prompt (e.g. "a car yields to a pedestrian") and processed through: Scene Planner → GMM Sampling → Waypoint Filter → Event Reasoner → TrajNet++ Refinement.

Three splits cover normal traffic, vehicle-pedestrian yielding, and collision scenarios. event_time_frame and metrics are only available for scenes that have simulation output logs (~40–72% per split). event_location_xy is only set for collision events; it is always null in the yield split.

Files / Splits

File	Scenes	Description
normal.jsonl	101	Normal background traffic, no scripted rare event
yield.jsonl	320	V2P: vehicle yields to pedestrian
collision.jsonl	621	V2V / V2P collision scenes
Total	1042

Fields

Field	Type	Description
scene_uid	str	Unique scene ID (dataset-sceneId-dirhash)
scene_id	str	Original scene ID (e.g. v2p_051)
category	str	normal / yield / collision
backend	str	LLM used for generation: claude3.5-sonnet / qwen2.5-7b / qwen2.5-3b / minimax / n/a
scene_type	str	v2p (vehicle–pedestrian) / v2v (vehicle–vehicle) / n/a
user_prompt	str	Natural language prompt given to the LLM
num_agents	int	Number of agents in the scene
relation_type	str	yield_to / collide_at / null
event_time_frame	int | null	Frame index of the primary Yield or Collision event; null if no output log
event_location_xy	[float, float] | null	[x, y] in metres; only set for Collision events
agents[]	list	Per-agent: id, type, goal, spawn_time, trajectory (T×2 float, metres)
events[]	list	Social-graph relation events from the scene definition
metrics	dict | null	collision_rate, min_distance_to_others, stop_before_crosswalk_success; null if no output log

Backend distribution

{
  "normal": {
    "n/a": 101
  },
  "yield": {
    "claude3.5-sonnet": 130,
    "minimax": 50,
    "qwen2.5-7b": 90,
    "qwen2.5-3b": 50
  },
  "collision": {
    "claude3.5-sonnet": 101,
    "minimax": 400,
    "qwen2.5-7b": 80,
    "qwen2.5-3b": 40
  }
}

Usage

from datasets import load_dataset

ds = load_dataset("JojoZhu/cosmos-trajectory-synthesis", data_files={
    "normal":    "normal.jsonl",
    "yield":     "yield.jsonl",
    "collision": "collision.jsonl",
})

# filter by backend
yield_claude = ds["yield"].filter(lambda x: x["backend"] == "claude3.5-sonnet")