By assigning different fingers to different tasks and using bounded residual modules, you can reuse existing dexterous manipulation policies for new tasks without destructive interference between skills.
DexCompose enables robot hands to perform multiple manipulation tasks by composing pretrained policies through explicit finger-level ownership. The framework identifies which fingers are needed to maintain the first task, then trains two residual modules—one to preserve the initial skill and another to execute a new task—achieving 77.4% success on composite manipulation tasks.