TAVIS: A Benchmark for Egocentric Active Vision and Anticipatory Gaze in Imitation Learning

Researchers introduced TAVIS, a comprehensive benchmark for evaluating active vision in imitation learning systems where robotic policies control their own gaze during manipulation tasks. The benchmark includes evaluation protocols, a novel metric (GALT) measuring anticipatory gaze, and baseline experiments showing that active vision benefits are task-dependent rather than universally beneficial.

TAVIS addresses a critical gap in robotics research by establishing the first standardized benchmark for active vision in imitation learning. As robotics systems increasingly adopt human-like gaze control mechanisms, the absence of comparative evaluation frameworks has hindered progress. This benchmark enables researchers to quantify when and how active vision contributes to manipulation success across different task types and embodiments.

The research builds on growing evidence that policies controlling their own visual focus—mimicking human attention mechanisms—improve learning efficiency. However, prior work lacked systematic evaluation, making it unclear whether benefits were universal or context-specific. TAVIS addresses this through paired experiments comparing headcam-controlled systems against fixed-camera baselines on identical demonstrations, providing controlled ablation studies.

The introduction of GALT (Gaze-Action Lead Time) represents a notable contribution bridging cognitive science and robotics. Grounded in human factors research, this metric quantifies how far in advance learned policies anticipate visual information needs, directly measuring if imitation learning captures human-like predictive gaze behavior. Initial results show median lead times comparable to human teleoperators, suggesting imitation learning naturally captures anticipatory attention patterns without explicit supervision.

The findings reveal task-conditional benefits and sharp performance degradation under distribution shifts, providing actionable insights for practitioners. Multi-task policies showed particular vulnerability, indicating that scaling active vision systems requires addressing robustness challenges. The public release of 2,200 demonstrations, evaluation code, and trained baselines democratizes access to high-quality robotics research infrastructure, accelerating the field's progress on a previously unmeasured problem.

• →TAVIS provides the first standardized benchmark for comparing active vision approaches in imitation learning across multiple robotic embodiments
• →Active vision benefits are task-conditional rather than uniformly helpful, requiring context-specific evaluation for deployment decisions
• →GALT metric quantifies anticipatory gaze in learned policies, finding imitation alone captures human-like predictive attention without explicit supervision
• →Multi-task policies degrade sharply under distribution shifts, highlighting a critical robustness challenge for scaling active vision systems
• →Publicly released infrastructure including 2,200 episodes and trained baselines enables reproducible research and accelerates active vision adoption