No physics engine, no renderer: MIRA simulates Rocket League from raw video
MIRA is the team's "world model" (software that learns to simulate an environment by watching video), and they are pointing it at physical AI rather than at games.
MIRA is the team's "world model" (software that learns to simulate an environment by watching video), and they are pointing it at physical AI rather than at games.
The MIRA team's world model learns to simulate Rocket League directly from raw video, with no physics engine, no renderer, and no hand-coded game state anywhere in the loop. In plain terms, the system watches a four-player 2v2 match, frame by frame, and learns to predict what comes next. According to the project blog, it runs that match at roughly 20 frames per second in 576p resolution on a single GPU, and stays stable across sessions long enough to play multiple goals.
The architecture pairs a 5-billion-parameter diffusion transformer with a 600-million-parameter video representation codec. The codec compresses each frame into a compact token representation, the transformer predicts the next chunk of tokens, and the codec decodes those tokens back into pixels. The model never sees a physics solver, a 3D scene graph, or a boost-meter variable in code. It learns ball physics, player cars, demolitions, and the in-game event text like "Shot on goal" or "Save" directly from the visual stream.
Long-horizon stability has been a chronic failure mode for causal video models. A frame-by-frame predictor whose small errors compound can drift out of distribution within seconds, which is why most learn-to-simulate demos either loop on a few seconds of footage or fall apart once the camera angle or the speed changes. The MIRA team frames sustained stable play across a full match as the genuine technical contribution, with the car-football demo as the visible payoff.
The autopilot is the second piece. The transformer is trained with action dropout, where it is occasionally forced to condition on the next player's intended action rather than its own free prediction. The team describes this as a way to drive the simulation in autopilot mode by letting the learned model "imagine" how a goalie or an attacker would behave, with action labels supplied by Nexto, an open-source Rocket League bot. The point is a control surface: a predictor that takes a real action as input is also one that can take a learned action as input, which is what makes the autopilot possible.
Both pieces earn the lab's positioning. The project page and GitHub repo call MIRA a "stepping stone to physical AI," where data is messier and scarcer than Rocket League footage and where there is no engine to fall back on. The lab is betting that if a transformer can learn the physics of a fast, multiplayer, physics-heavy game from pixels alone, the same training recipe could later back simulators for robots learning from real-world video. The team is also clear that MIRA is not being used to develop Rocket League itself, with the README and the GitHub repo both stating that explicitly.
That framing sets the watch items. The hardware spec that MIRA runs on a single NVIDIA B200 comes from AI Weekly's release summary, not from the lab's own post, so a reader should treat that as aggregator-reported until a primary-source check confirms it. The preprint on arXiv is not peer-reviewed. Discussion on Hacker News is a single submitter with three points, which is community signal rather than independent reproduction.
The thing to watch is whether the open dataset, training and inference code under mira-wm/mira gets picked up by an outside lab and whether MIRA's stability gets quantified in a benchmark an outside group can rerun. The physical-AI claim is a direction, not a result, but a direction with code attached is one a peer lab can actually falsify.