TrajReasoner

TrajReasoner is a visual navigation dataset for fine-tuning vision-language models to reason about terrain traversability and output robot trajectories directly from camera observations.

Each example contains an RGB camera frame, projected future trajectory supervision, and robot pose metadata. The trajectory masks provide visual supervision, while uv_center_json provides the projected trajectory centerline as image coordinates.

Dataset Structure

Each row represents one processed camera frame.

Column	Type	Description
id	string	Stable row id in the format <bag>_<frame_idx>
bag	string	Source ROS bag stem
frame_idx	int32	Frame index within the processed bag
image	Image	RGB camera frame resized to 320x240
trajectory_mask	Image	Binary filled trajectory ribbon mask
centerline_mask	Image	Binary center-line trajectory mask
depth_image	Image	Normalized depth visualization when available
position	list[float32]	Robot world-frame XY position
yaw	float32	Robot heading in radians
uv_center_json	string	JSON-encoded projected center trajectory points

Usage

from datasets import load_dataset

ds = load_dataset("gianluca-capezzuto/trajreasoner", split="train")
sample = ds[0]

image = sample["image"]
trajectory_points = sample["uv_center_json"]
trajectory_mask = sample["trajectory_mask"]

Notes

• RGB frames and masks are stored at 320x240.
• Mask images use 255 for trajectory pixels and 0 for background.
• Invalid or missing depth pixels are black in depth_image.
• uv_center_json may be an empty list for frames with no valid projected future trajectory.