Skill Acquisition via Automated Multi-Coordinate Cost Balancing

Harish Ravichandar, Seyed Reza Ahmadzadeh, Muhammad Asif Rana, Sonia Chernova

Abstract

Learning from Demonstration (LfD) typically relies on a single coordinate system, often Cartesian, which may not capture the essential geometric features of diverse tasks. We present Multi-Coordinate Cost Balancing (MCCB), a framework that automatically integrates multiple differential coordinate systems—Cartesian, Tangent, and Laplacian. By solving a meta-optimization problem to weight these coordinates based on demonstration variance, MCCB achieves superior generalization and accuracy across various robotic tasks compared to traditional single-coordinate methods.

Introduction

In robotic Learning from Demonstration (LfD), the choice of coordinate system significantly impacts the fidelity and adaptability of skill reproduction. While Cartesian coordinates are standard, they often fail to capture local shape features (Laplacian) or directional intent (Tangent) critical for specific tasks like drawing or pushing. Identifying the optimal coordinate system for an unknown skill is a non-trivial challenge for human programmers.

Comparison of trajectory reproductions using different coordinate systems versus the MCCB framework.

As shown in Figure 1, relying on a single coordinate system can lead to poor generalization when starting or ending points deviate from demonstrations. MCCB addresses this by simultaneously analyzing demonstrations in multiple spaces and automatically determining their relative importance through a data-driven approach.

The MCCB Framework

The MCCB framework operates through a structured pipeline that transforms raw demonstrations into a robust, multi-coordinate cost function for trajectory generation.

The MCCB pipeline: from multi-coordinate encoding and joint density estimation to automated cost balancing and convex optimization.

1. Multi-Coordinate Encoding

Demonstrations are projected into three distinct differential coordinate systems:

  • Cartesian: Captures absolute spatial positions.
  • Tangent: Represents first-order differences, emphasizing the direction of motion.
  • Laplacian: Represents second-order differences, capturing local curvature and shape.

Each coordinate set is modeled using a Gaussian Mixture Model (GMM) to represent the joint density of the state and time, capturing both the mean trajectory and its associated variance at every time step.

2. Blended Cost Function and Automated Balancing

MCCB defines a global cost function that penalizes deviations across all three coordinate spaces. The core innovation lies in the Automated Cost Balancing stage. Instead of manual tuning, the system solves a meta-optimization problem to calculate weighting factors ($\alpha$) that minimize reproduction error relative to the original demonstrations. This process effectively identifies which geometric features are "critical" (low variance) versus "flexible" (high variance), adhering to the minimal intervention principle.

3. Constrained Optimization

The final reproduction is generated by solving a convex optimization problem. This ensures the robot satisfies physical constraints—such as specific starting points, targets, and via-points—while minimizing the learned multi-coordinate cost.

Experimental Evaluation

The framework was validated on the LASA Handwriting Dataset (25 2D motions) and three 3D robotic tasks: picking magnets from blocks, pressing pegs into holes, and pushing a box lid closed.

Key Results

MCCB consistently outperformed single-coordinate baselines and uniform weighting schemes across metrics like Sum of Squared Errors (SSE) and Frechet Distance.

  • Task-Specific Adaptation: MCCB correctly identified the dominant coordinate for each task: Laplacian for pushing ($\alpha_L = 0.94$), Cartesian for pressing ($\alpha_C = 0.67$), and Tangent for handwriting.
  • Error Reduction: In the pushing task, the Cartesian-only baseline produced an error nearly six times higher than MCCB.
  • Generalization: The system matched or exceeded state-of-the-art performance on handwriting benchmarks without requiring manual parameter adjustment.

Conclusion

Multi-Coordinate Cost Balancing (MCCB) provides a robust solution to the coordinate selection problem in LfD. By automatically weighting Cartesian, Tangent, and Laplacian features, the framework enables robots to learn and reproduce complex skills with high precision and adaptability. Future work will explore expanding the coordinate pool and incorporating dynamic environment constraints into the optimization phase.

Summary updated

Abstract

We propose a learning framework, named Multi-Coordinate Cost Balancing (MCCB), to address the problem of acquiring point-to-point movement skills from demonstrations. MCCB encodes demonstrations simultaneously in multiple differential coordinates that specify local geometric properties. MCCB generates reproductions by solving a convex optimization problem with a multi-coordinate cost function and linear constraints on the reproductions, such as initial, target, and via points. Further, since the relative importance of each coordinate system in the cost function might be unknown for a given skill, MCCB learns optimal weighting factors that balance the cost function. We demonstrate the effectiveness of MCCB via detailed experiments conducted on one handwriting dataset and three complex skill datasets.
BibTeX
@inproceedings{ravichandar2019skill,
  author = {Harish Ravichandar and Seyed Reza Ahmadzadeh and Muhammad Asif Rana and Sonia Chernova},
  title = {{Skill Acquisition via Automated Multi-Coordinate Cost Balancing}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=zKehLWcbIG},
}
Figure