This paper introduces Curriculum-Guided Active Learning (C-GAL), a framework that optimizes demonstration selection for compositional robotic skills. C-GAL uses factorized skill embeddings and ensemble-based uncertainty to identify high-value demonstrations, while a curriculum scheduler increases task difficulty progressively. The method achieves significant improvements in sample efficiency and generalization on a benchmark with 256 compositional combinations.
Key findings
C-GAL achieves 32% improvement in sample efficiency over random selection.
The framework reaches an 82.4% success rate with only 5,000 demonstrations.
C-GAL shows 18.3% better generalization to held-out combinations compared to random selection.
The approach explores 92% of the compositional skill space.
Limitations & open questions
The framework's performance may vary with different numbers and types of factors in compositional tasks.
The scalability of C-GAL to real-world robotic tasks with higher complexity is yet to be tested.