General AI Challenge










Challenge Update #13 - important information for the participants: your input needed!

June 1, 2017

1st Round evaluation & your submitted solutions

Please take two minutes to let us know what type of solution you plan to submit, and the type of HW and OS you plan to use. Open the survey here.


We are hard at work preparing the evaluation for the first round. We would like to ask you a couple of questions to have a better estimate of what to prepare for. We’d also like to share our plans so far. This description is just a preliminary plan. You will get detailed instructions soon enough.

The evaluation will run on one of two types of Azure VMs, either CPU-bound (with no GPU) or GPU-bound. We plan to run Linux-based solutions in Docker (nVidia Docker in case of GPU-bound VMs). The container will have an open port for access to the CommAI environment, and no access to the internet. The images will be self-contained without a need for external data.

On Windows the situation is different since nVidia Docker is not available there. Solutions using GPU will not be run in Docker. We have not yet decided about the CPU-bound Windows-based solutions. We will specify the details after we evaluate the survey linked above.


the deadline for your EGPAI submission is getting closer!


The EGPAI 2017  (Evaluating General-Purpose AI workshop, August 19-25, 2017, Melbourne, Australia) will host a special session for the General AI Challenge track, reporting on the state of the competition, including an invited talk given by GoodAI and a few selected short reports from the participants.

Registered participants are invited to submit a short summary (2-4 pages) explaining their approach to solving the challenge and their experience so far. The deadline is June 10th, 2017.

IMPORTANT: If you plan to submit to EGPAI, please let us know today by sending a quick note to

Reminder: 1st Round specifications document update

In case you would like to refresh your memory and check out how will we be evaluating your solutions, have a look at the 1st Round Specifications document. Note that it underwent an update in March. In case you missed the update, be sure to read the 2nd version of the document on the Challenge website or by following the link above.

We would also like to ask you not to send us the data you used to train your agent as part of your submission, due to its possibly large size. It would be enough if you send us your pre-trained agent. In the case we need to have a look at the data, we will reach out to participants individually. We will also recap what needs to be part of your submission closer to the deadline. In case you have any questions, please let us know at

News from the AI world

Automatic Goal Generation for Reinforcement Learning Agents

by David Held, Xinyang Geng, Carlos Florensa, Pieter Abbeel


"We use a generator network to propose tasks for the agent to try to achieve, specified as goal states. The generator network is optimized using adversarial training to produce tasks that are always at the appropriate level of difficulty for the agent. Our method thus automatically produces a curriculum of tasks for the agent to learn."

Safeguarding the future: how to outsmart the AI?


Watch the recording from the Globsec panel with Marek Rosa (CEO/CTO of GoodAI and the founder of the General AI Challenge), Khalil Rouhana (Director for "Digital Industry" in DG CONNECT, European Commission), and Andreas Ebert (EU Technology Officer, Microsoft).

Everything that Works Works Because it's Bayesian: Why Deep Nets Generalize?

"HINT: because they are really just an approximation to Bayesian machine learning."

Please reload

  • Facebook - White Circle
  • Twitter - White Circle
  • YouTube - White Circle
  • 49398-reddit-social-logo-character-icon-vector-icon-vector-eps
  • Blogger - White Circle