Skip to main content

OpenAI's Dota 2 bots remove 'significant' match restrictions, move closer to human-level skill

Last month, we wrote about how OpenAI's Dota 2 bots are totally on some I, Robot shit as they aim to "beat a team of top professionals" at The International. Back then, the collective known as the OpenAI Five were working to set of tight restrictions. Now, a host of the "most significant" conditions have been removed. And it's getting exciting. 

It's worth noting here that I know next to nothing about Dota 2. I'm now going to unashamedly pass off PCG dep ed/MOBA expert Pip's thoughts and advice on the following as my own.  

So, it appears OpenAI has now reintroduced Roshan—the river-dwelling NPC monster which bisects the map—as fair game. Killing him grants victors significant bonuses, therefore an important part of any match is knowing when your enemy is trying to kill him—or judging when best to attempt it yourself. OpenAI now reckons the OpenAI Five are smart enough to handle these considerations on the fly. 

Likewise, wards have been reinstated. Wards expand players' ground vision—which would have previously given human players a leg-up over their AI counterparts—and are therefore tied to Roshan, as knowing when your enemies are targeting the river baddie is crucial mid-battle.   

"Because our training system Rapid is very general, we were able to teach OpenAI Five many complex skills since June simply by integrating new features and randomizations," explains OpenAI in a blog post. "Many people pointed out that wards and Roshan were particularly important to include—and now we’ve done so. We’ve also increased the hero pool to 18 heroes. Many commenters thought these improvements would take another year. We’ll see how well these new game mechanics work on August 5th."

OpenAI says that while the bots' strength comes more from teamwork and coordination than reflexes, it has managed to up the reaction time of the Five from 80ms to 200ms—a reaction time closer to human level. 

As OpenAI notes above, this next match—the 'OpenAI Five Benchmark match'—takes place on August 5, ahead of the Five's final showdown at The International 2018. 

"The OpenAI Five Benchmark match will be held 12:30pm Pacific Time on August 5th in San Francisco," adds OpenAI. "The human team will include Blitz, Cap, Fogged, and Merlini, some of whom are former professionals." 

NB—Thanks, Pip!