Event

CapPro: Adapting VLA to New Tasks from a Single Human Video

1 reports · 1 sources

Researchers Qi Chen et al. propose the CapPro framework, which enables vision-language-action (VLA) policies to adapt to previously unseen manipulation tasks from just a single human demonstration video, without requiring robot demonstrations of the target task and without fine-tuning for the target task. The framework consists of a phase-aware human video representation module and a video-conditioned adaptation module: the former learns task-relevant query representations through visual token selection and phase description supervision; the latter injects the compact video and query representations into the action-conditioning context of a pretrained VLA policy. The authors state that the human demonstration provides task execution guidance rather than a robot-embodiment-aligned trajectory for imitation. The authors report that experiments on source tasks and unseen target tasks show that CapPro improves average performance on unseen target groups while maintaining comparable performance on source tasks.