LAMP Dataset

arXiv | Webpage | Code | Model Weights

This dataset accompanies the CVPR 2026 paper LAMP: Language-Assisted Motion Planning. For complete instructions on the overall framework, model inference, and environment setup, please refer to Code.

Dataset Overview

The dataset consists of a single JSON file containing 200,000 training entries. It is formatted as a conversational exchange between a user (human) and the model (gpt) to facilitate instruction-tuning for motion planning tasks.

Structure

Each entry in the JSON array represents a single conversational sequence. The conversation strictly covers three instruction-following tasks in the following order:

• Object Motion: A natural language description of an object's movement, prompting the model to output the expected Box Trajectory.
• Camera Motion: A natural language description of the camera's behavior, prompting the model to output the expected Camera Tags.
• Camera Editing: A prompt requesting a modification to the previously generated camera movement, prompting the model to output the updated Edit Tags.

Example Entry

[
  {
    "id": "traj_74062",
    "conversations": [
      {
        "from": "human",
        "value": "After an abrupt descent, the advanced drone dramatically zipped inward, delicately adjusting its path with a subtle leftward drift through several moments, then concluded with a sweeping flight to the left. Generate Box Trajectory: "
      },
      {
        "from": "gpt",
        "value": "free-form t_x: no t_y: far_down t_z: no yaw: 0 tilt: 0 roll: 0 t_x: no t_y: no t_z: far_in yaw: -20 tilt: 0 roll: 0 t_x: no t_y: no t_z: far_in yaw: -20 tilt: 0 roll: 0 t_x: far_left t_y: no t_z: no yaw: 0 tilt: 0 roll: 0"
      },
      {
        "from": "human",
        "value": "A dramatic optical push-in from a static camera position brings the speeding car's insignia into sharp focus, followed by a responsive rotation to track its trajectory. Generate Camera Tags: "
      },
      {
        "from": "gpt",
        "value": "rotation_track push_in_0.5"
      },
      {
        "from": "human",
        "value": "rotation_track push_in_0.5 Slow down the push/pull and start retreating. Generate Edit Tags: "
      },
      {
        "from": "gpt",
        "value": "rotation_track push_in_0.3 goes_out_1_0.5"
      }
    ]
  }
]

Camera Parameters

To obtain the 6-DoF camera parameters, tag-to-trajectory conversion is needed. Please see the Code for the conversion scripts, as well as training and inference.

Citation

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

@misc{kizil2025lamp,
    title={LAMP: Language-Assisted Motion Planning for Controllable Video Generation}, 
    author={Muhammed Burak Kizil and Enes Sanli and Niloy J. Mitra and Erkut Erdem and Aykut Erdem and Duygu Ceylan},
    year={2025},
    eprint={2512.03619},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={[https://arxiv.org/abs/2512.03619](https://arxiv.org/abs/2512.03619)}, 
}