SAVR: Scooping Adaptation for Variable food properties via Reinforcement Learning

Abstract

Personalizing bite sizes is crucial for robot-assisted feeding, as users have diverse dietary needs and preferences. However, precisely controlling the amount of food scooped remains a challenge due to variations in food properties, such as texture, granularity and cohesion. This work introduces SAVR (Scooping Adaptation for Variable food properties via Reinforcement learning), a learning-based framework that enables robots to scoop a targeted amount of food while adapting to different food characteristics. SAVR integrates Dynamic Motion Primitives (DMPs), with Reinforcement Learning (RL), where DMPs provide a structured motion representation, and RL refines execution by modifying the force term within the DMP formulation. This formulation enables efficient learning by allowing the RL agent to fine-tune the scooping trajectory rather than learning entire trajectories from scratch. Through ablation studies, we show that segmented spoon and food masks, combined with force-torque data, are essential for accurate scooping, significantly improving sim-to-real transfer. We validate SAVR on a real robotic system, demonstrating substantial improvements in accuracy and adaptability over baselines, without any additional fine-tuning.