Synthetic multi-agent traffic trajectories at a fixed urban intersection, generated by the COSMOS pipeline with Protocol V2 LLM backends and Tier-1 quality gates (geometry + kinematics).

Each scene is created from a natural language prompt and processed through: Scene Planner → GMM Sampling → Waypoint Filter → Event Reasoner → TrajNet++ Refinement.

Trajectories in agents[].trajectory are the refined (final) stage, not GMM originals.

This release has no normal-traffic split; only rare-event yield and collision cohorts. The same scene_id may appear up to four times (one per LLM backend)—use scene_uid or backend + scene_id as the stable key.

event_time_frame and metrics are present when simulation artifacts exist under the synthesis output tree. event_location_xy is only set for collision-style primary events.

Organization

The dataset is structured as follows:

.
├── collision.jsonl    # V2V collision scenes (JSON Lines, one scene per line)
├── yield.jsonl        # V2P yield scenes
├── stats.json         # Scene counts and backend distribution
└── README.md

Each line in collision.jsonl / yield.jsonl is one JSON object (scene). See Fields below.

Files / Splits

File	Scenes	Description
`yield.jsonl`	400	V2P: vehicle yields to pedestrian
`collision.jsonl`	400	V2V collision scenes
Total	800

Fields

Field	Type	Description
`scene_uid`	str	Unique ID (`dataset-sceneId-hash(pool_key)`)
`scene_id`	str	Scene index within interaction type (e.g. `v2v_051`)
`category`	str	`yield` / `collision`
`backend`	str	Report name, e.g. `claude-sonnet-4.6`
`backend_key`	str	Internal Protocol V2 alias, e.g. `claude_sonnet_4_6`
`scene_type`	str	`v2p` or `v2v`
`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 primary Yield/Collision event
`event_location_xy`	[float, float] \| null	`[x, y]` in metres for collision events
`agents[]`	list	`id`, `type`, `goal`, `spawn_time`, `trajectory` (T×2 m, refined)
`events[]`	list	Social-graph relation events
`metrics`	dict \| null	Simulation metrics when available
`pool_key`	str	Protocol V2 Tier-1 manifest key (reproducibility)

Backend distribution

{
  "yield": {
    "claude-sonnet-4.5": 100,
    "claude-sonnet-4.6": 100,
    "deepseek-v4-flash": 100,
    "minimax-m2.7": 100
  },
  "collision": {
    "claude-sonnet-4.5": 100,
    "claude-sonnet-4.6": 100,
    "deepseek-v4-flash": 100,
    "minimax-m2.7": 100
  }
}

Usage

from datasets import load_dataset

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

# unique scene (backend-specific)
row = ds["collision"].filter(lambda x: x["scene_uid"] == "...")[0]

Citation

If you use this dataset, please cite the following papers:

@misc{zang2026language,
  author       = {Zang, Chengbo and Zhu, Wangshu and Turkcan, Mehmet Kerem and Zussman, Gil and Ghaderi, Javad and Kostic, Zoran},
  title        = {Language-Guided Traffic Simulation and Rare Event Synthesis for an Urban Intersection},
  howpublished = {CVPR Workshop},
  year         = {2026}
}

@inproceedings{raychaudhuri2020cosmos,
  author    = {Raychaudhuri, Dipankar and Seskar, Ivan and Zussman, Gil and Korakis, Thanasis and Kilper, Dan and Chen, Tingjun and Kolodziejski, Jakub and Sherman, Michael and Kostic, Zoran and Gu, Xiaoxiong and Krishnaswamy, Harish and Maheshwari, Sumit and Skrimponis, Panagiotis and Gutterman, Craig},
  title     = {Challenge: COSMOS: A city-scale programmable testbed for experimentation with advanced wireless},
  year      = {2020},
  isbn      = {9781450370851},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  doi       = {10.1145/3372224.3380891},
  booktitle = {Proceedings of the 26th Annual International Conference on Mobile Computing and Networking},
  articleno = {14},
  numpages  = {13},
  location  = {London, United Kingdom},
  series    = {MobiCom '20}
}

Acknowledgements

This work was supported in part by NSF grants CNS1827923 and EEC-2133516, NSF grant CNS-2038984 and corresponding support from the Federal Highway Administration (FHA), NSF grant CNS-2148128 and by funds from federal agency and industry partners as specified in the Resilient & Intelligent NextG Systems (RINGS) program, and ARO grant W911NF2210031.

Downloads last month: 75

Total file size:

13.4 MB