We’ve been talking about adding RL to our bots lately to scale, which leads to the idea:
The RL training loop that powers your SC2 bot is architecturally the same one Boston Dynamics uses to train Atlas.
It’s a simple, portable engine: observe → act → reward → learn.
Whether the environment is a physics simulator or a game engine, the loop doesn’t change. You’re building a system that solves for a goal in a complex environment.
The Transferable Skill
You’ve spent hours debugging reward hacking and tuning hyperparameters. I’ve been there, staring at a bot that found a weird loophole to get points without actually winning.
You are practicing the exact skill that trips up robotics engineers daily. OpenAI proved this when they took the exact same reinforcement learning pipeline that beat the world champions at Dota 2, and pointed it at a physical robotic hand. The game was just the cheapest, fastest place to simulate failure at scale.
Packaging the Signal
Tell a small, short story. Make it fun. Turn your bot’s achievement into content fuel.
Boston Dynamics could have just posted a whitepaper. Instead, they showed us Atlas failing, adjusting, and finally nailing the kick. Then they tied it directly to the real world: how the balance learned in the game translates to moving through a warehouse.
You can do the exact same thing. Post a short GIF of your bot failing, then succeeding. Tell a quick story, and explicitly name the real-world problem your pipeline could actually be solving.
