Abstract
As robots transition from industrial settings to human-centric environments like homes and offices, they must navigate complex social norms. A fundamental social construct is ownership, which dictates permissions and prohibitions regarding object interaction. This paper presents a robotic system that learns ownership relations and associated norms through a combination of direct human instruction and probabilistic inference. By integrating symbolic logic with machine learning, the system allows robots to represent, learn, and infer these social rules in real-time, even under conditions of perceptual uncertainty.
Introduction
For a robot to be truly useful in a social environment, it needs to understand more than just geometry and physics; it must grasp social "norms." Ownership is a deceptively complex norm that implies a set of permissions. For instance, a robot should distinguish between discarding an empty soda can and a colleague's unopened drink. Current systems often struggle with this because social rules are frequently unwritten, context-dependent, and learned through interaction. This research addresses how a robot can effectively represent and infer ownership relations and the norms governing them.
Proposed Methodology
The researchers developed a hybrid system combining three distinct approaches to manage social norms:
- Predicate-Based Norm Representation: Norms are represented using first-order logic (e.g.,
forbid trash if ownedBy any). This provides an explicit and interpretable framework for the robot's rule-following behavior. - Incremental Rule Learning: The system employs a "dual-mode instruction" capability. It can learn a general rule from a single direct command (one-shot learning) or induce a general rule by observing several specific examples (inductive generalization).
- Probabilistic Ownership Inference:
- Perception: The robot utilizes Kernel Logistic Regression (KLR) to predict ownership based on 3D position, color, and recent human interaction.
- Bayesian Reasoning: The system uses Bayesian inference to "back-calculate" ownership likelihood. If a robot is forbidden from an action (e.g., throwing an item away), it concludes the object is likely owned based on the existing norm "don't throw away owned things."
Experimental Evaluation
The system was evaluated in both simulated and real-world environments:
- Simulated Environment: A custom simulation with 20 colored blocks and 3 agents was used for large-scale testing (100 trials) to measure rule induction accuracy under noisy conditions.
- Real-World Scenarios: The system was deployed on a Baxter research robot. Scenarios included responding to verbal reprimands, refusing commands that would violate a user's ownership rights, and inducing general norms from specific permissions regarding objects on a table.
Key Results and Discussion
The experimental results demonstrated the system's effectiveness in learning and applying social norms:
- Robustness to Noise: In simulations, the norm-learning component achieved over 76% accuracy in predicting permissions, even when ownership data was noisy or incomplete. This highlights the strength of the probabilistic inference model.
- Error Reduction: In workspace clearing tasks, the robot committed significantly fewer mistakes (averaging ~6.3 errors) compared to a baseline robot that assumed all actions were permitted (which resulted in 8.75 to 15 errors).
- Generalization and Inference: The robot successfully generalized commands (e.g., applying a "don't touch" rule from one object to others owned by the same user) and correctly identified social relations from normative feedback—treating a "No!" not just as a stop command, but as a data point about an object's social status.
Conclusion and Future Work
This research represents a significant step toward robots that can autonomously navigate human social structures. However, several limitations remain. Current norms are "self-centric," meaning the robot only learns rules for its own actions rather than observing norms between humans. Future work will explore temporal complexities like borrowing, handling conflicting norms (e.g., safety vs. ownership), and incorporating more complex social cues into the perceptual model.
Summary updated