rainbowrobotics/open_door_mobile_all_260819

RBY1 Open Door Mobile Dataset

This dataset contains mobile manipulation demonstrations for approaching and opening a door using the Rainbow Robotics RB-Y1 robot.

The dataset was created in the LeRobot v3.0 format.

Dataset Summary

Item	Value
Robot	Rainbow Robotics RB-Y1
Format	LeRobot v3.0
Episodes	154
Frames	143,620
FPS	30
Duration	Approximately 1 hour 19 minutes 47 seconds
Tasks	2
Cameras	Front and right
Action dimension	23
State dimension	44

Task Instructions

The task instruction depends on the initial position of the robot.

Task instruction	Description	Episodes	Frames
Open the door	The robot starts near the door and focuses on opening it.	49	25,866
Approach the door and open it	The robot approaches the door, stops in front of it, and opens it.	105	117,754

Source Datasets

Final episodes	Source dataset	Task instruction
0–48	rainbowrobotics/open_door_mobile_turning_260807	Open the door
49–80	rainbowrobotics/open_door_mobile_turning_260812	Approach the door and open it
81–153	rainbowrobotics/open_door_mobile_turning_260819	Approach the door and open it

Feature Schema

Feature	Type	Shape
observation.images.front	Video	480 × 640 × 3
observation.images.right	Video	640 × 480 × 3
observation.state	float32	44
action	float32	23
task_index	int64	1
timestamp	float32	1
episode_index	int64	1
frame_index	int64	1

Loading the Dataset

from lerobot.datasets import LeRobotDataset

dataset = LeRobotDataset(
    "rainbowrobotics/open_door_mobile_all_260819",
    revision="main",
)

print(dataset)
print(dataset[0]["task"])

PI0.5 Training

The natural-language instructions are stored in the dataset task metadata.

When training PI0.5 with OpenPI, enable task prompts using:

base_config=DataConfig(prompt_from_task=True)

Dataset Processing

The three source datasets were merged without video re-encoding. Feature schemas, FPS, episode ordering, episode lengths, frame counts, and per-frame task indices were verified after merging.

Episode 22 of open_door_mobile_turning_260819 was trimmed by 25 seconds before merging. It corresponds to final episode 152 and contains 915 frames.