Driving · 2026-10-01

DashVMC: Real-Time Discrete World Model Control in Geometry Dash

Florent Tariolle, Florian Yger

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

DashVMC learns a compact, action-conditioned discrete world model from ~2 hours of recorded Geometry Dash gameplay, then improves a BC-initialized controller using PPO entirely in frozen-model rollouts (no live-game interaction). The refined policies survive longer than BC across three controller seeds on all three official levels and a held-out community layout. A deployment baseline skips visual generation and sustains a 60-Hz capture-to-action loop on an RTX 2060.

Why it matters

The paper addresses the mismatch between imagination-based policy improvement and live, non-paused control where capture→prediction→action must fit within frame deadlines. DashVMC evaluates whether offline learned dynamics can be actionable for control, and shows both improved live survival (vs BC) and an engineered inference path meeting a 60-Hz loop constraint.

Method

Learns a compact discrete, action-conditioned world model from approximately two hours of recorded Geometry Dash, predicting a 64-token next grid (parallel) plus an alive/death score.
Initializes an action model with behavioural cloning from recorded actions, then refines it with PPO using imagined rollouts from the frozen dynamics model (no further live-game interaction).
Deploys a reactive control loop that skips visual generation and uses the transformer’s temporal summary in a 60-Hz capture-to-action loop.

Limitation

DashVMC studies one deterministic binary-action game, with one post-freeze layout within known mechanics; repeated level prefixes make model diagnostics in-distribution. Three controller seeds share one tokenizer and dynamics checkpoint, measuring controller rather than end-to-end uncertainty, and death calibration covers only recorded states, leaving recursive reward exploitation unquantified.

Abstract (from arXiv)

World-model agents are usually evaluated in simulators that can wait for the policy; live games impose the opposite constraint, requiring capture, prediction, and action before the next frame. We present DashVMC, which learns a compact, action-conditioned world model from approximately two hours of recorded Geometry Dash gameplay. To test whether the learned dynamics are actionable, a controller is initialized by behavioural cloning (BC) and refined with Proximal Policy Optimization (PPO) entirely in frozen-model rollouts, without further interaction with the live game. Across three controller seeds, the refined policies survive longer than their BC initializations on all three official levels and a held-out community layout. At deployment, the baseline skips visual generation and sustains a 60-Hz capture-to-action loop on a consumer GPU. Action-conditioned continuations and rollout diagnostics show that the model remains useful for control despite imperfect long-horizon fidelity.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-09-30 · Indexed by Pub-AI: 2026-10-01 UTC

preprint