Abstract
This paper presents a novel framework for robotic manipulation that enables autonomous identification of task-relevant object features. By leveraging meta-level priors—learned relationships between geometric features and task relevance from previous skills—the system significantly improves feature selection accuracy. Experimental results on a PR2 robot across six diverse manipulation tasks demonstrate that this approach allows for robust generalization with minimal demonstrations, achieving a 77.8% success rate in novel scenarios.
Introduction
When a robot learns a skill from a human demonstration, it is often overwhelmed by a high-dimensional feature space. For a task like pouring, the robot must distinguish between critical features (e.g., the cup's rim) and irrelevant ones (e.g., the table's color). Traditionally, this "needle in a haystack" problem is solved by manual feature engineering or large-scale statistical learning. This paper addresses the challenge of learning feature relevance autonomously and efficiently by transferring knowledge from previously learned skills.
Methodology
The proposed framework treats feature selection as a transfer learning problem, utilizing "meta-level priors" to guide the robot's attention. The system architecture is divided into several key stages:
- Part-Based Feature Generation: Objects are segmented into functional parts using the GrabCut algorithm, focusing on areas of interaction during demonstrations.
- Meta-Features and Priors: The core innovation involves calculating "meta-features" that describe the geometric relationship between a feature and the task (e.g., proximity to the hand, alignment with motion). The robot learns a mapping from these meta-features to the probability of relevance, creating a meta-level prior.
- Feature Selection and Execution: The system uses Stochastic Search Variable Selection (SSVS), a Bayesian approach, to identify a sparse set of relevant features. This process is initialized with the meta-level prior rather than a uniform distribution. Finally, the motion is executed using Dynamic Movement Primitives (DMPs) adapted to the selected features.
Experimental Setup
The approach was evaluated using a PR2 robot performing six manipulation tasks: placing, pushing, tilting, cutting, pouring, and wiping. The researchers collected 90 demonstrations across three different object sets, primarily utilizing the YCB object set— a standard benchmark for household manipulation. The goal was to test if a robot could generalize these skills to novel objects and configurations with only 15 demonstrations per task.
Results and Discussion
The inclusion of meta-level priors led to significant performance gains:
- Accuracy: The use of meta-priors more than doubled the average precision and recall in feature selection compared to standard uniform priors.
- Generalization: The robot showed a superior ability to predict correct goal states in new environments. On the physical PR2 platform, the system achieved a 77.8% success rate across various tasks in novel situations.
- Transfer Logic: A key insight is that feature relevance is grounded in physical geometry. By encoding these relationships, a robot that has learned to "cut" can apply the logic of what makes a part relevant to a new task like "pouring."
Conclusion
By transferring the logic of how to learn relevance rather than just transferring trajectories, this research enables robots to become more autonomous and data-efficient. While the system faced challenges with specific tasks like "pushing" where priors were occasionally misleading, the overall framework demonstrates a powerful path toward robots that can quickly adapt to complex, unstructured human environments. Future work will focus on handling variable object counts and integrating real-time sensory feedback.
Summary updated