Flow Control: Steering Vision-Language-Action Models with Simple Real-Time Inputs
Flow control introduces a simple method to steer vision-language-action (VLA) models in real-time using generic inputs like keyboard commands, without needing retraining. It injects user intent into the VLA's flow matching action expert by modifying the initial condition of the ODE. This allows crude user inputs to guide complex robot actions toward desired outcomes, significantly increasing task success rates.