Abstract
A successful robot-assisted feeding system requires bite acquisition of a wide variety of food items. It must adapt to changing user food preferences under uncertain visual and physical environments. Different food items in different environmental conditions require different manipulation strategies for successful bite acquisition. Therefore, a key challenge is how to handle previously unseen food items with very different success rate distributions over strategy. Combining low-level controllers and planners into discrete action trajectories, we show that the problem can be represented using a linear contextual bandit setting. We construct a simulated environment using a doubly robust loss estimate from previously seen food items, which we use to tune the parameters of off-the-shelf contextual bandit algorithms. Finally, we demonstrate empirically on a robot-assisted feeding system that, even starting with a model trained on thousands of skewering attempts on dissimilar previously seen food items, ε-greedy and LinUCB algorithms can quickly converge to the most successful manipulation strategy.
Introduction
Robotic-assisted feeding is a critical technology for millions of people with mobility impairments. However, "bite acquisition"—the act of successfully picking up food with a utensil—is surprisingly difficult. Food items vary wildly in texture, shape, and consistency. A strategy that works for a grape might fail for a slice of banana.
Existing systems often rely on "batch-trained" models, such as SPANet, which learn from large, static datasets. A significant hurdle for these models is covariate shift: when a robot encounters a food item or environment (lighting, plate type) not represented in its training data, its performance plummets. As shown in Figure 1, there is a marked decrease in success rate when a model encounters "Out-Of-Class" items compared to those it was trained on.
Methodology
To address this, the authors frame the feeding task as a Linear Contextual Bandit problem, allowing the system to learn "on the fly" from its own successes and failures.
The Framework
- Context ($x$): The robot captures an RGBD image of the food, which is processed through the penultimate layer of a pre-trained SPANet to extract a high-dimensional feature vector.
- Actions ($a$): The robot selects from six discrete skewering strategies. These are defined by fork pitch—Vertical (VS), Tilted-Vertical (TV), or Tilted-Angle (TA)—and roll (Parallel or Perpendicular to the food's long axis). These actions are visualized in Figure 2.
- Reward: A binary signal where success is defined as the food remaining on the fork for five seconds.
Algorithms
The study compared two classic bandit approaches:
- ε-greedy: Primarily selects the best-known action but explores randomly with probability ε.
- LinUCB (Linear Upper Confidence Bound): Selects actions based on both estimated success rates and the uncertainty of those estimates, naturally favoring the exploration of unknown strategies.
To tune these algorithms safely, the researchers utilized a "doubly robust" estimator to repurpose offline data into a simulated training environment.
Experimental Results
The system was evaluated on a physical ADA (Autonomous Dexterous Arm) using food items excluded from the initial training set, such as bananas and carrots.
- Rapid Convergence: In real-robot experiments, the algorithms typically identified the optimal skewering strategy within approximately 10 failures per new food item.
- Robustness of Features: A major finding was that features learned from one set of foods (e.g., carrots) are descriptive enough to help a bandit algorithm learn about a completely different food (e.g., a banana) very quickly.
- Stability with LinUCB: While ε-greedy is simpler, LinUCB proved more robust. It avoids getting "stuck" after early fluke failures because it explicitly tracks the uncertainty of its predictions.
- Adaptation: Even when starting with a model that performed worse than random on a specific food, the online framework quickly reached a success rate comparable to a fully supervised model (~75%).
Conclusion and Future Work
Online learning is not merely an enhancement but a necessity for robust robotic feeding. By adapting to unseen items in real-time, the system overcomes the limitations of static datasets.
Future research will aim to expand the discrete action space into continuous variables (force, speed, complex angles) and address "cluttered" plates where food items overlap. Additionally, integrating haptic feedback—the "feel" of the food—could further improve performance on visually similar but texturally distinct items, such as ripe versus unripe fruit.
Summary updated