SeeTraceAct: Visibility-Aware Latent Planning from Cross-Embodiment Demonstration Videos
Researchers have developed SeeTraceAct, a new framework for robot policies that can learn from a single demonstration video. This approach addresses limitations in existing models that struggle with precise localization of small targets. SeeTraceAct improves performance by predicting future end-effector traces with visibility awareness. The framework was tested on a new dataset, RoboCasa-DC, and a real-world benchmark, showing significant improvements in success rates compared to existing methods. AI
IMPACT This research could enable robots to learn new tasks more efficiently from limited demonstrations, potentially accelerating their deployment in complex environments.