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

The researchers proposed a scalable diffusion strategy called SCRIPT, aimed at achieving language-driven physical humanoid robot control through a multi-stage training framework. The core of this approach is the Joint Action-State-Text Diffusion Transformer (JAST-DiT), which represents actions, physical states, and text as dedicated token streams and couples them through a joint attention mechanism. To stabilize the autoregressive control process, the system introduces a nonlinear historical condition mechanism to retain recent dense context and sample long-term sparse cues. In the training process, in addition to supervised imitation pre-training, a reinforcement learning post-training phase (RLHR) with combined rewards is designed. By injecting learnable noise into the flow sampling process and utilizing mixed physical feedback and text rewards in a closed-loop simulation environment, the quality of motion and command compliance are improved. Quantitative evaluations show that SCRIPT outperforms existing state-of-the-art methods, and extended studies on the MotionMillion dataset demonstrate consistent performance improvement with model size increase. The relevant code will be made available for…