The ShapeStacks dataset.
A simulation dataset for vision-based intuitive physics.
Python interface and Tensorflow models.
Examples for data handling and model training on ShapeStacks with Python and Tensorflow.
Tech report.
Learning vision-based physical intuition for generalised object stacking.
Dataset
20,000 simulation scenarios
RGB, segmentation and depth modalities
Detailed annotation of (in)stability
Source Code
Data loader for Tensorflow dataset API
Script examples for stability prediction
Pre-trained stability predictor
Tech Report
Dataset construction & details
Intuitive physics of stability prediction
Application to object stacking
ShapeStacks Dataset
ShapeStacks-Manual.md
MuJoCo world definitions (39 MB):
shapestacks-mjcf.tar.gz
·
shapestacks-mjcf.md5
Meta information (156 KB):
shapestacks-meta.tar.gz
·
shapestacks-meta.md5
RGB images (33 GB):
shapestacks-rgb.tar.gz
·
shapestacks-rgb.md5
Violation segmentation maps (875 MB):
shapestacks-vseg.tar.gz
·
shapestacks-vseg.md5
Depth maps (1.1 GB):
shapestacks-depth.tar.gz
·
shapestacks-depth.md5
Source Code
Handling of the ShapeStacks data with Tensorflow dataset API.
Example scripts for training and evaluation of stablity prediction models.
Check out the code on GitHub »
Pre-trained stability prediction models based on InceptionV4 (1.8 GB):
shapestacks-incpv4.tar.gz
·
shapestacks-incpv4.md5
Tech Report
Physical intuition is pivotal for intelligent agents to perform complex tasks.
In this paper we investigate the passive acquisition of an intuitive understanding
of physical principles as well as the active utilisation of this intuition in the
context of generalised object stacking. To this end, we provide ShapeStacks: a
simulation-based dataset featuring 20,000 stack configurations composed of a variety
of elementary geometric primitives richly annotated regarding semantics and structural
stability.
[paper]
·
[poster]
·
[video]
Research funded by