SCRIPT: Scalable Diffusion Policy with Multi-stage Training for Language-driven Physics-Based Humanoid Control

SCRIPT is a new approach for controlling physics-based humanoids from natural-language instructions, aimed at closing the gap between expressive command following and physically believable motion. The system uses a Joint Action-State-Text Diffusion Transformer, or JAST-DiT, to keep language, body state, and action tokens in the same attention space so the controller can reason about semantics and dynamics together.

The training pipeline is split into stages. First comes supervised imitation pre-training, then a post-training phase called Reinforcement Learning with Hybrid Rewards. That second stage mixes physical feedback with text-based rewards and injects learnable noise into flow sampling, which is meant to improve both motion quality and instruction adherence during closed-loop simulation.

The practical angle for game developers is obvious: this is the kind of research that could feed future character controllers, embodied NPCs, and animation systems that need to react to free-form prompts without collapsing into unstable motion. The authors say SCRIPT outperforms prior methods on text alignment, motion quality, and physical realism, and that scaling continues to help as the model grows.

A notable detail is the use of the 1200-hour MotionMillion dataset, where scaling studies showed consistent gains. The code is planned for public release, which should make it easier for researchers and engine teams to test whether diffusion-based control can be adapted to production-facing character pipelines.