pbelevich/lego42176_garage_parking

LEGO 42176 Garage Parking Dataset

Imitation learning dataset for driving a LEGO Technic 42176 car into a garage. Recorded with LeRobot v3.0 format.

A human operator drives the car via keyboard over Bluetooth while an IP camera records the scene. The policy learns to map camera observations to speed/steering commands.

Dataset summary

Episodes	100
Total frames	12,037
Duration	13.4 minutes
FPS	15
Resolution	640 x 480
Robot type	lego42176
Task	Drive the LEGO car into the garage

Episode statistics

	Min	Median	Mean	Max
Frames	48	113	120	260
Duration (s)	3.2	7.5	8.0	17.3

Features

Feature	Type	Shape	Description
observation.image	video (h264)	[480, 640, 3]	Camera view of the car and garage
observation.state	float32	[2]	Current [speed, steering] normalized to [-1, 1]
action	float32	[2]	Commanded [speed, steering] normalized to [-1, 1]

Action space

Actions are continuous values in [-0.8, 0.8] (normalized from the LEGO hub's [-80, 80] range):

• speed: positive = forward, negative = reverse
• steering: positive = right, negative = left

How it was recorded

1. LEGO 42176 car connected via Bluetooth Low Energy (BLE)
2. IP camera streaming RTSP at 640x480
3. Human drives the car from a starting position into a garage using keyboard (arrow keys)
4. Each episode = one parking attempt (press R to start, drive, press R to save)
5. Dataset finalized after each episode to prevent data loss

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("pbelevich/lego42176_garage_parking")

sample = dataset[0]
print(sample["observation.image"].shape)  # torch.Size([3, 480, 640])
print(sample["action"].shape)             # torch.Size([2])

Training

This dataset was used to train an ACT (Action Chunking with Transformers) policy on Apple Silicon (MPS backend). See the project repository for training and inference scripts.

Hardware

• Car: LEGO Technic 42176 Porsche GT4 e-Performance
• Camera: Dahua IP camera, 640x480 via RTSP (substream)
• Compute: MacBook Pro M1 Pro