Please select the desired project time frame:

• July 2026
• January 2026
• July 2025
• January 2025
• July 2024
• January 2024
• July 2023
• January 2023
• July 2022
• January 2022
• July 2021
• January 2021
• July 2020
• January 2020
• July 2019
• January 2019
• July 2018
• January 2018
• July 2017
• January 2017
• July 2016
• January 2016
• July 2015
• January 2015
• July 2014
• January 2014
• July 2013
• January 2013
• July 2012
• January 2012
• July 2011
• January 2011
• July 2010
• January 2010
• July 2009
• January 2009
• July 2008
• January 2008
• July 2007
• January 2007
• July 2006
• January 2006
• July 2005
• January 2005
• July 2004
• January 2004
• July 2003
• January 2003
• July 2002
• January 2002
• July 2001
• January 2001

Start of funding 01.01.2019

Instrinsically-motivated Intelligent Systems - Unsupervised Learning through Physical Interaction and Observation

---

Infants explore their world in a seemingly arbitrary but intrinsically structured way, commonly described as curiosity. Attracted by coherences they haven’t yet understood, babies seem to intentionally create events that are new, informative and exciting to them. An artificial system that could mimic such abilities would be of great use for applications in computer vision, robotics, reinforcement learning, and many other areas. This research aims at further investigation of possibilities for an autonomous intelligent system to learn without supervision and self-supervise in realistic physical environment. In the scope of this project, we are combining computer science techniques, e.g deep learning, with the psychology successes to make yet another step towards better artificial intelligence. The results are expected to be initial steps toward creating flexible self-supervised autonomous agents and further develop efficient learning methods. This project benefits from partners that come from various areas of research, Computer Science, Neuroscience and Psychology.

Final report:
In this BaCaTec project we focused on a joint research with Stanford university. Recently, neural-network based forward dynamics models have been proposed that attempt to learn the dynamics of physical systems in a deterministic way. While near-term motion can be predicted accurately, long-term predictions suffer from accumulating input and prediction errors which can lead to plausible but different trajectories that diverge from the ground truth. A system that predicts distributions of the future physical states for long time horizons based on its uncertainty is thus a promising solution. In this project, we introduce a novel robust Monte Carlo sampling based graph-convolutional dropout method that allows us to sample multiple plausible trajectories for an initial state given a neural-network based forward dynamics predictor. By introducing a new shape preservation loss and training our dynamics model recurrently, we stabilize long-term predictions. We show that our model's long-term forward dynamics prediction errors on complicated physical interactions of rigid and deformable objects of various shapes are significantly lower than existing strong baselines. Lastly, we demonstrate how generating multiple trajectories with our Monte Carlo dropout method can be used to train model-free reinforcement learning agents faster and to better solutions on simple manipulation tasks.