Reward-Free Scooter Balance Control via Diffusion World Models with Goal-Conditioned Trajectory Generation

We present a reward-free control framework for balancing and steering a two-wheeled scooter using a diffusion-based world model. Rather than engineering a reward, we specify goals directly in observation space: target values (e.g., zero roll and zero yaw error) are pinned through a continuous mask, and classifier-free guidance amplifies the goal signal during trajectory generation. Because the mask is continuous at inference, goals can be traded off online (for instance, relaxing the balance constraint during sharp turns to allow necessary leaning) without retraining. The model is a FiLM-Mixer denoising network trained with V-prediction diffusion. At deployment, the controller runs in real time using a single diffusion step with warm-started predictions. We validate the approach on a full-sized Thormang3 humanoid operating a Gogoro Viva scooter in simulation, and deploy it on physical hardware. It matches a PPO baseline tuned with six reward components on balance, survival, and heading tracking while producing smoother commands, all without the per-task reward-shaping step. Diffusion training introduces its own loss-weight hyperparameters; unlike reward weights, however, these are task-agnostic. They govern the denoising procedure rather than the desired behavior, and are therefore set once and reused unchanged across goals rather than re-tuned for each new task. Because the model learns to predict trajectories rather than to maximize a reward, its training signal depends only on observed states and actions, not on reward labels. Real hardware recordings can therefore be folded directly into the same loss, providing a route toward closing the sim-to-real gap that reward-based methods such as PPO structurally cannot use.