GIFT: Generalizing Intent for Flexible Test-Time Rewards

A team from UC Berkeley (Fin Amin, Nathaniel Dennler, Andreea Bobu) has introduced GIFT (Generalizing Intent for Flexible Test-Time Rewards), a novel framework tackling a core robotics problem: reward functions learned from demonstrations often fail when a robot encounters new environments. The issue is that current methods latch onto spurious visual or semantic correlations in the training data, rather than the underlying human intent. GIFT solves this by leveraging language models to infer high-level intent by contrasting preferred demonstrations with non-preferred ones, grounding reward generalization in what the human actually cares about.

At deployment, GIFT maps novel test states to behaviorally equivalent training states using this inferred intent as a conditioning signal. This allows a robot's learned reward function to generalize to new situations—like different objects or table layouts—without requiring any retraining. The team evaluated GIFT on tabletop manipulation tasks, testing it across four simulated tasks with over 50 unseen objects. It consistently outperformed visual and semantic similarity baselines in key metrics like test-time pairwise win rate and state-alignment F1 score.

The research was validated in the real world with a 7-Degree-of-Freedom Franka Panda robot, demonstrating reliable physical transfer. The work, set to appear at the IEEE International Conference on Robotics and Automation (ICRA) 2026, represents a significant step toward more adaptable and robust robotic systems that can understand and execute tasks based on human goals, not just surface-level cues.