Bayesian Optimization for Developmental Robotics with Meta-Learning by Parameters Bounds Reduction

Maxime Petit, Emmanuel Dellandréa, Liming Chen

Introduction

In industrial robotics, the "bin-picking" task—where a robot must identify and grasp specific objects from a cluttered container—remains a significant challenge. Achieving high success rates often requires fine-tuning numerous software hyperparameters. Traditionally, this is done manually or through Bayesian Optimization (BO). However, standard BO faces a "cold-start" problem: it treats every new object as a completely novel task, ignoring prior knowledge gained from similar items. This lack of "life experience" makes optimization expensive, time-consuming, and prone to noise in real-world environments.

This paper introduces a developmental robotics framework that leverages meta-learning and long-term memory to "warm-start" the optimization process. By identifying visual similarities between objects and reducing the search space for hyperparameters, the system significantly accelerates learning and improves grasping reliability.

Proposed Framework and Methodology

The core of the proposed system is a cognitive architecture that integrates long-term memory with specialized reasoning modules to guide the optimization process.

The proposed developmental robotics architecture integrating long-term memory with reasoning modules for parameter optimization.

Cognitive Architecture

The system utilizes three distinct memory types:

  1. Episodic Memory: Stores raw data and scores from every individual optimization trial.
  2. Procedural Memory: Maintains the best-performing parameter sets and the resulting reduced search bounds for specific objects.
  3. Semantic Memory: Contains 3D CAD models and visual features of known objects.

Parameters Bounds Reduction

The primary innovation is the Parameters Bounds Reduction module. When a robot encounters a new object, it uses a Visual Similarity Module (powered by PointNet) to find the most similar "ancestor" in its semantic memory. Instead of starting a global search, the robot analyzes the distribution of successful parameters for that ancestor. Using statistical tests (Dudewicz-van der Meulen for uniformity and Wilcoxon for medians), the system identifies high-potential ranges for each of the 9 hyperparameters and shrinks the search bounds accordingly. This forces the Bayesian Optimization—specifically using the Expected Quantile Improvement (EQI) criterion—to focus its limited budget on the most promising regions of the search space.

Experimental Evaluation

The framework was tested in both simulation and real-world environments:

  • Simulation: Evaluated using 8 diverse objects, including industrial components and household items from the Princeton Shape Benchmark.
  • Real-World: Validated using an industrial Fanuc robot setup tasked with grasping elbowed rubber tubes.
  • Optimization Task: The system optimized 9 continuous hyperparameters within the "Kamido" professional grasping software suite.

Key Results and Discussion

The researchers imposed a strict optimization budget of only 30 iterations per object, a challenging constraint for a 9-dimensional search space.

Comparison of grasping success rates over 30 iterations: standard Bayesian Optimization (left) vs. the proposed meta-learning approach with parameter bounds reduction (right).

Performance Gains

As shown in the performance comparison, the meta-learning approach provided a clear "warm-start" advantage. During the initial design phase (the first 10 trials), the meta-learning robot achieved approximately 75% success, while the cold-start robot hovered around 65%. By the end of the 30 iterations, the meta-learning framework reached an overall success rate of 84.3%, significantly outperforming the standard BO rate of 78.9% (p-value = 0.036).

Reliability and Industrial Impact

Beyond peak success rates, the framework drastically improved worst-case performance. Minimum success rates jumped from 28.3% to 70.6%, suggesting that the system is far more reliable for industrial deployment where consistency is paramount. In real-world physical tests, the mean performance improved from 75.6% to 84.6%.

Conclusion

This research demonstrates that "learning" in robotics is as much about knowing where not to look as it is about finding the optimal value. By treating the search space as a form of transferable knowledge, the robot can leverage its past experiences to master new tasks with minimal trials. Future work aims to expand this by allowing the system to dynamically expand search bounds if necessary and linking these mathematical bounds to human-readable symbolic labels.

Summary updated

BibTeX
@inproceedings{petit2020bayesian,
  author = {Maxime Petit and Emmanuel Dellandréa and Liming Chen},
  title = {{Bayesian Optimization for Developmental Robotics with Meta-Learning by Parameters Bounds Reduction}},
  booktitle = {ICDL-EPIROB},
  year = {2020},
  doi = {10.1109/icdl-epirob48136.2020.9278071},
}
Figure