All blogs / Coder One 2022 Recap: Bomberland heads outer space and Igglysplat reigns supreme
September 26, 2022 • Joy Zhang • Updates • 7 minutes
Coder One (formerly, the AI Sports Challenge) is an AI programming tournament that took place between 1-14 September 2022. The challenge involves building an AI capable of autonomously competing in a multiplayer game, played entirely by other AI. Teams from across the globe competed in our fourth season for a slice of the $5,000AUD prize pool and a feature in our casted livestream finale.
After an intense 12 days of programming and over 80k matches played, we crowned team Igglysplat as our Coder One grand champions of 2022. Keep reading to hear more about how the competition developed, what we learned, and what's next.
Bomberland is a multi-agent adversarial environment inspired by the classic console game, Bomberman. Teams control three units that compete in a 2D grid-world collecting power-ups and placing explosives. The goal is to be the last remaining team alive.
In its fourth season the Bomberland environment received an upgrade including: a new space theme, freeze powerup, ammo system, and mechanic for spawning powerups.
These changes were made to encourage more head-to-head action, discourage a 'stand in the centre tile at all costs'-strategy, and reward efficient block farming (an easy way to measure progress in the early stages of development).
We were excited to see much more action and a significantly higher bar for strong agents compared to previous seasons!
If you’re interested in checking out the environment, Bomberland is available open source for anyone to use. We welcome suggestions and contributions!
With your feedback, we were able to make various improvements to our competition platform. Here’s what got shipped this season:
As part of our journey to make Coder One the most exciting AI programming tournament on the planet, this season we experimented with a new Circuit Point tournament format.
At the end of each day, teams received a fixed number of Circuit Points (CP) based on their ranking on the leaderboard. This was intended to incentivize early submissions, and disincentive teams from holding onto their submissions until the very last day.
As a result, we were excited to see teams submitting regularly throughout the competition and the meta evolving over time. From your feedback, we learned:
There were some interesting discussions on this topic — look forward to an improved system for the next competition based on your feedback!
We've summarised your additional feedback in a later section of this post.
We welcomed both new and returning teams from a wide range of backgrounds and countries. On the final day of the competition, a mad rush of last minute submissions saw 30k matches generated that took 6 hours to complete. In the aftermath, we celebrated our Top 4 Finalist teams:
Igglysplat (CP: 657)
Our season 1 champion and season 2 2nd-place winners.
Eop (CP: 610)
Eop is our Season 3 winner and an Android developer interested in AI and the future of humanity.
ryandy (CP: 531)
A new competitor to Coder One. Ryandy is an embedded software engineer who sacrificed some beauty sleep to make an impressive climb from 10th → 1st place on the leaderboard in the last moments of the competition.
bruh (CP: 521)
Team bruh finally takes a well-deserved Top 4 spot in their 3rd year of competing.
Many other teams made for strong contenders for the Top 4 spots but ultimately were pushed out in the final days of competition. A shout-out goes to:
Donkie2.0 (CP: 507)
Donkie is back! Donkie is a finalist from both Season 1 and Season 2. Donkie’s agent is coded in Python with only one mission: to take over Mars.
SigmaGo (CP: 506)
A freshly-graduated master of computer science from Croatia. Likes long walks on the beach and being obliterated in AI competitions online (apparently).
SigmaGo’s agent has one mission: to minimize.
The question: what?
The answer: we’ll figure that out later.
NotABot (CP: 295)
A team of Computer Science Undergraduates from IIT Indore, India who started out learning about machine learning and various other algorithms before coming across this competition. NotABot’s agent tries to capture the centre and then collect powerups and bomb the opponents.
ColomboKnights (CP: 209)
Just a regular guy from Sri Lanka pretending to be a team of Knights and trying to navigate the exciting world of reinforcement learning. ColomboKnights' agent deploys a simple and straightforward strategy:
if empty_surrounding_tiles: move() else: self_destruct()
Here’s a recap of the top moments from stream:
The first set of the finale featured some bruh moments which were nonetheless entertaining.
More than once, these teams showed us that being down 1v3 units didn’t necessarily spell defeat:
This was an extremely close nailbiter of a set that ultimately resulted in a 3-2 win for Igglysplat.
We saw ryandy’s agent for the first time on our live finale stage. While ryandy’s agent was particular efficient at collecting powerups and being aggressive, Eop’s strong map control and centre hogging strategy proved dominant.
This was yet another extremely close set, with Eop finally taking the 3-2 win against ryandy.
The final stage was a dramatic reunion with the Season 1 winner meeting the Season 3 winner for the first time.
But ultimately, it was our Season 1 champions Igglysplat returning to the stage in spectacular fashion to take home Season 4.
👏👏👏 Congratulations to Igglysplat and the finalist teams: Eop, ryandy, and bruh 👏👏👏 You can find the final weighted leaderboard results for Season 4 here.
Looking to revisit the finale moments? The finale VOD is now available on our Twitch channel.
Here’s a quick summary of what we’re hearing from you:
Thanks to everyone who shared feedback with us! Your feedback will go a long way in helping us improve the next competition.
You can link also your open-source agents and write-ups or check out other community contributions from the Bomberland repo.