Introduction
In Human-Robot Interaction (HRI), robots typically acquire skills through demonstrations or reinforcement learning. However, these learned skills are often "brittle." For example, if a robot learns a recipe using specific ingredients, switching to an alternative usually requires the robot to either relearn the entire task or undergo extensive retraining. Sub-symbolic representations, such as neural network weights, are particularly difficult to adjust because modifying a single parameter to fix one action can inadvertently degrade other unrelated skills.
This research addresses the lack of flexibility in current robotic systems, proposing a framework where a human can provide simple natural language instructions—such as "When you do X, do Y instead of Z"—to update the robot's internal logic instantly without losing prior knowledge.
The Skill Modification Framework
The proposed framework is built upon two primary pillars designed to move away from "black box" models toward transparent, modifiable logic:
- Action Scripts: These are symbolic representations of tasks. Each script includes a signature (arguments), pre-conditions (requirements to start), operating conditions (requirements during execution), and post-conditions (the expected result).
- Formal Operators: The system maps natural language to three mathematical operators:
- $O_{ins}$ (Insert): Adds a new step or condition.
- $O_{del}$ (Delete): Removes a step or condition.
- $O_{rep}$ (Replace): Swaps one component for another.
These operators allow for recursive modification, enabling users to tweak high-level scripts or specific sub-steps. The framework was implemented within the DIARC (Distributed Integrated Affect Reflection Cognition) architecture, allowing the robot to translate spoken English into these formal operations on the fly.
Experimental Evaluation
The framework was evaluated through three distinct lenses:
- Algorithmic Verification: A synthetic test of 50 randomly generated action scripts. The authors performed 450 automated modifications, achieving a 100% success rate in maintaining logical consistency.
- HRI Case Studies: Real-world validation using a PR2 and a Nao robot. Tasks included passing objects and performing physical routines like squats.
- User Study: A study involving 31 participants who provided feedback on robot failure corrections. Results indicated that 67% of users preferred natural language modification over graphical interfaces or direct programming.
Results and Discussion
The framework demonstrated significant versatility in real-world scenarios, successfully performing four types of modifications:
- Generalizing: Adapting a "pass" action to work for any person rather than a fixed target.
- Repairing: Correcting a routine by adding missing steps (e.g., lowering arms after a squat).
- Generating: Creating a new "show" skill by modifying an existing "pass" skill.
- Adding Effects: Teaching the robot new post-conditions, such as the fact that picking up an object results in "holding" it.
Key insights from the study suggest that transparency is vital for effective correction. When a robot can narrate its "thought process," humans can more easily pinpoint errors. Furthermore, participants reported higher levels of trust when they felt empowered to modify the robot's behavior directly.
Conclusion and Future Work
By treating natural language as a "compiler" for symbolic action scripts, this framework enables immediate, transparent skill modification. While highly effective, future work must address the "ripple effect," where modifying a sub-skill might impact other tasks that rely on it. Additionally, scaling the system to handle the inherent ambiguity of free-form natural language and ensuring physical safety during real-time logic updates remain critical areas for further research.
Summary updated