Improbable Icon


A match made in... someplace amazing?



So, Unity announced yesterday the initial preview release of ml-agents. I’ve started digging into this (since I spend a fairly significant amount of time on ml “things” for my day job as well, I’m extremely curious about their approach to things like PPO optimization).

As it stands right now, ml-agents is a python platform that works “with” Unity (through some ostensible voodoo that I don’t yet understand); and since there’s no python sdk for spatial, it’s unclear how this would work. But the concept of a solid RL implementation paired with the scalability of the spatial platform and the ease of using unity literally gives me goosebumps in the potential to team these things up for simulations.

I’m curious what you guys’ thoughts on this are – this seems like a great opportunity for Improbable to team up with the Unity AI folks and create an incredible platform on from my perspective.


After reading the Unity blog post, I’m excited as well!

It seems that the “academy” is what need to be shared, so it should be on the SpatialOS layer of things.
Workers containing brains and agents.

Skynet here we come!


I just had a conversation with a co-worker around the possibility of Automotive AI built on a simulation driven by adversarial RL. Imagine training your automotive AI in a simulation with hundreds of different car models, where agents are driving against each other. Such an approach could provide that missing component of simulating “driver’s intuition” (in the same way that AlphaGO demonstrated such capabilities). With spatial, one could feasibly simulate hundreds of car “entity templates” all with different capabilities; along with modeling environmental conditions, and simulating various system failures in the cars sensors - then add spatial’s scaling to create an environment with 100’s of thousands of vehicles in the world all feeding into both a local brain and a shared brain?

Literally… Goosebumps… Skynet is right, my friend. I mean - the benefits of a persistent scalable simulation upon which to train AI in this way is exciting and frightening on so many levels.

Another interesting possibility - from the game dev perspective is the ability to use AI for “playtesting” your games; which is always a tricky endeavor (to say the least). Kicking off a few thousand AI agents as new characters in your MMO and allowing them to train individually, form teams around character strengths and weaknesses, etc. RL by itself is a very exciting field, but putting this in the hands of Unity devs? I’ll be surprised if we don’t start seeing some amazing progress made around gen purpose ai as a result of it’s accessibility to small studios.


Yeah I read their post about this the other day! I was excited to try to use the economy driven machine learning so my brains that I’m making now by myself would be able to learn, but they don’t have that example ready yet. Gives me more time to work on my brain code and get it ready in the meantime though!


Hey! this is exactly the sort of technology i’m interested in too.

I’m building something that allows people to create and compete AI agents in online environments.

wrote a blog post about it here:

The blog is a bit old… so the tech has moved on a bit… you can find it at

I’m looking for collaborators and co-conspirators on building this into something amazing.

Spatial is one of the technologies i’m currently looking into to build bigger environments then i current have. (single box tech).