FAQ
Table of Contents
- What is BabyBench?
- What is MIMo?
- Who can participate?
- I don’t have experience with reinforcement learning, can I still participate?
- Are there any baseline models or starter code?
- Are there restrictions on the algorithms or architectures I can use?
- Can I modify the BabyBench environments?
- What kind of submission is expected?
- How will the winner be chosen?
- What do I get if I win?
- Can I publish work based on my BabyBench submission?
- More questions?
What is BabyBench?
BabyBench is a multimodal benchmark of infant behaviors for developmental artificial intelligence. The BabyBench Competition hosted at IEEE ICDL invites participants to model infant-like learning in simulated environments using MIMo.
What is MIMo?
MIMo, the multimodal infant model, is a simulation platform based on the MuJoCo physics engine. You can find out more about MIMo here. You can also download and use MIMo for your own experiments here.
Who can participate?
Anyone! Students, researchers, interdisciplinary teams—no matter your background, you’re welcome. If you’re interested in developmental learning, this is for you.
I don’t have experience with reinforcement learning, can I still participate?
Absolutely, but training MIMo will require some basic knowledge of reinforcement learning. We recommend reading the resources section or heading to the discussion page on github to connect with others.
Are there any baseline models or starter code?
Yes! We provide some basic examples with starter code to help you get up and running with training MIMo using reinforcement learning here.
Are there restrictions on the algorithms or architectures I can use?
No, but we strongly encourage models that follow the spirit of the competition: unsupervised, self-supervised, or intrinsically motivated learning approaches that aim to recreate or explain developmental behaviors. Take into account that plausibility is one of the evaluated criteria. Having said that, creativity and interpretability are valued!
Can I modify the BabyBench environments?
In short, no. The environments are designed to facilitate the emergence of specific developmental behaviors, and the core challenge of the competition lies in enabling the agent to learn these behaviors autonomously through interaction and experience. The focus of the competition is on the learning algorithms, and winning team will be the one that achieves the best results given the constraints and limitations of the environments. However, we understand that some algorithms may require modifications to the environment, in which case submissions will need to justify the need for such modifications focusing on the biological plausibility of the learning mechanism.
What kind of submission is expected?
You’ll be asked to a 2-page abstract describing your approach and the log files automatically generated during training and evaluation. Details can be found here.
How will the winner be chosen?
Each submission will be scored out of 10 points, using a mixture of qualitative and quantitative criteria explained here.
What do I get if I win?
The winning team of the BabyBench Competition will receive a €150 prize. All finalists will receive a certificate at the IEEE ICDL conference.
Can I publish work based on my BabyBench submission?
Yes, absolutely. Your BabyBench work can be a great foundation for a contribution to a conference or journal. We are working on a full version of the BabyBench platform in the fall of 2025. We’re happy to discuss follow-up opportunities!
More questions?
We’re happy to help! Contact us by email or leave a message in the discussion page on github.