Abstract
Traditional Learning from Demonstration (LfD) often requires numerous examples to build generalized statistical models, which can be data-intensive and rigid when faced with novel environments. This paper proposes a Case-Based Reasoning (CBR) framework that treats individual demonstrations as unique "cases." By storing structured representations of tasks and employing a multi-level adaptation pipeline, a robot can successfully transfer knowledge from a single demonstration to a target environment with different object layouts and configurations.
Introduction
In interactive robotics, Learning from Demonstration (LfD) is a primary method for teaching agents new tasks. However, standard approaches typically rely on large datasets to create generalized models. This presents two significant challenges: high data requirements for human teachers and a lack of robustness when the robot encounters environments that deviate from the training distribution.
To address these issues, we propose a framework based on Case-Based Reasoning. Instead of discarding individual experiences in favor of a mean representation, our approach preserves specific demonstrations as cases. This allows a robot to retrieve a single, relevant past experience and intelligently adapt it to a new "target" environment, such as a kitchen with different counter heights or object placements.
Methods
The proposed framework is divided into two primary phases: Source Case Storage and Case Adaptation for Target Environment.
Case Representation
A "case" is defined as a structured tuple $C = <L, D, T, O, S_i, S_f>$, where:
- L (Label): The task identifier (e.g., "scooping").
- D (Action Models): Motion encoded via Dynamic Movement Primitives (DMPs), allowing for scalable trajectories.
- T (Parameterization): Functions relating motion to specific objects.
- O (Salient Objects): Critical objects required for the task.
- $S_i, S_f$ (States): Initial and final environment snapshots.
The Adaptation Pipeline
When a robot is tasked with performing a known action in a new environment, it follows a five-step pipeline:
- Observation: Perceiving the new environment (e.g., via overhead camera).
- Retrieval: Identifying the most relevant stored case based on goals and object similarity.
- Mapping: Establishing correspondences between objects in the source case and the target environment.
- Transfer: Adjusting the stored motion trajectory to fit new object coordinates.
- Execution: Performing the adapted plan.
A key feature of this framework is multi-level abstraction. The robot can vary the level of detail transferred based on environmental similarity. If the target environment is highly similar, low-level details are preserved; if it is significantly different, the robot relies on high-level task principles to maintain robustness.
Preliminary Results and Insights
Initial experiments conducted with a robotic agent focused on tabletop manipulations like scooping and pouring. The results demonstrate the feasibility of the Case Storage and Transfer/Execution processes.
Key insights include:
- Value of Individual Experiences: CBR allows every demonstration to serve as a distinct tool, mirroring human episodic memory.
- The Mapping Challenge: A critical component of successful transfer is the "mapping bridge"—correctly identifying functional roles of objects across different scenes.
- Robustness through Abstraction: By adjusting the level of abstraction during transfer, the robot can successfully utilize a single source demonstration to solve tasks in varied target environments.
Limitations and Future Work
Current limitations include the reliance on human intervention for object mapping and case retrieval. Future research will focus on automating these components, specifically developing autonomous object mapping algorithms and more sophisticated retrieval metrics to enhance the robot's ability to independently recall and adapt past experiences.
Summary updated