This one is pure childhood
Real Steel is one of those movies that basically wired my brain as a kid. Giant boxing robots, controlled by a human, fighting it out in a ring. It’s a big part of why I fell in love with robotics in the first place, so this project is about as me as it gets.
A while back I found a Roblox game built around Real Steel, and I got completely hooked. I actually climbed all the way to #6 on the global leaderboard (I still have the screenshots, I’m not letting anyone forget that one). But the more I played, the more one thought kept nagging at me: in the movie, the best fighters don’t just mash buttons, they command their robot. So what if I could stop pressing keys entirely and just fight with my voice, like Max does with Atom?
So I built exactly that.
What it does
It’s a Python program that listens to my microphone and turns spoken words into moves in the game, in real time. I say “right,” the robot throws a right. “Hook,” it hooks. And it goes further than single punches:
- Modifiers stack onto a punch. Say “heavy upper” and it throws a heavy uppercut. “Body left” throws a left to the body. Under the hood these hold a modifier key down, press the punch, then release.
- Named combos fire whole sequences. Shout “fire combo” and it rips off a scripted four-hit string; “destroyer” unleashes a five-hit finisher. These felt amazing to pull off by voice.
- Block holds the guard for you: say “block” and it presses and holds the block key for a couple of seconds on its own thread, then releases.
Everything is a single word on purpose, because in a real fight you do not have time for full sentences.
One dumb detail I love: the mic I shout into is my dad’s work headset, the one I end up borrowing almost every day. There’s something ridiculous and great about sitting there in a chunky work headset barking “destroyer” at my monitor, and yes, I’ve actually won a few matches doing exactly that.
The whole game was speed
The fun part and the hard part were the same thing: latency. A boxing match is fast. If the program waited for me to finish a sentence, recognise it, and then act, I’d already be knocked out. So the real engineering went into making it feel instant.
I used Vosk, an offline speech-recognition engine, so there was no round trip to a server slowing things down. But the real trick was that I don’t wait for a finished phrase at all, I act on partial results while I’m still talking. The moment Vosk thinks it heard “right,” the punch fires, before I’ve even finished the word.
That instant-fire approach creates its own mess, though. If you naively act on every partial update, a single “right” gets fired ten times as the recogniser keeps re-reporting it. So I had to build a little debounce system: each command remembers when it last fired, ignores repeats within a fraction of a second, and tracks which words it already handled in the current phrase. Tuning those cooldowns, low enough to spam punches when you mean to, high enough to not double-fire by accident, was most of the actual work.
BASE_COMMANDS = {'right': 'i', 'left': 'j', 'upper': 'l', 'hook': 'k', ...}
MODIFIERS = {'heavy': 'd', 'body': 's'}
# "heavy upper" -> hold D, press L, release D
There was one catch I didn’t see coming. To keep things fast I used a small, lightweight Vosk model, and a light model buys its speed by giving up accuracy. Mid-fight it would mishear me, or worse, mix up two commands that sounded too alike, so a punch I called wouldn’t land or the wrong one would. The fix turned out to be more about language than code: I renamed the commands to short, punchy, phonetically distinct words the model could nail even when I was rushing. “Jab” became “right,” “uppercut” became “upper,” and so on, picked as much for how cleanly they get recognised as for what they mean. Turns out designing a good voice interface is half engineering, half choosing words that don’t sound like each other.
Blocks and combos run on their own threads so that holding a guard or ripping a five-hit combo never freezes the ear that’s still listening for my next shout.
Why I love it
Is a voice-controlled boxing script practical? Absolutely not. Was it one of the most fun things I’ve built? Completely. It taught me real things too, real-time audio pipelines, working with a speech-recognition engine, injecting input into a game, and the surprisingly deep art of tuning latency so something feels instant.
But mostly, it let a kid who grew up obsessed with Real Steel actually stand in his room and shout a robot into throwing a fire combo. Some projects are for the CV. This one was for me.