Search papers, labs, and topics across Lattice.
This paper introduces TacEvo, a self-evolving architecture discovery framework that leverages large language models (LLMs) to enhance neural network designs for robotic tactile perception. By employing a MAP-Elites quality-diversity loop and generating code-level mutations, TacEvo significantly improves architectural diversity and efficiency, achieving a remarkable 56.1% increase in validation fitness over 20 generations. The framework not only matches expert performance in force prediction but also surpasses it in fine-grained grating classification, highlighting its potential for advancing AI-assisted scientific discovery in specialized robotic sensing.
LLM-driven self-evolving architectures can outperform expert-designed models in robotic tactile perception tasks, achieving unprecedented levels of performance and diversity.
Vision-based tactile sensing converts contact-induced surface deformation into images, enabling robots to infer contact forces and fine surface textures that are not accessible through conventional vision alone. However, tactile images are sensor- and physics-specific, so effective architectures often require expert intuition and extensive manual iteration. Existing neural architecture search (NAS) pipelines can reduce this burden, but they are often computationally expensive and restricted to hand-designed search spaces, which limits architectural novelty and diversity. We introduce TacEvo, a self-evolving architecture discovery framework that improves network designs from downstream feedback. TacEvo uses an LLM to generate code-level mutations and crossovers, and a MAP-Elites quality-diversity loop that preserves diverse elite architectures while preferentially reusing prompts that consistently yield improvements. Exploration is guided by two behavioural descriptors, Architectural Diversity and Efficiency Ratio, which encourage coverage across structural variations and compute-size trade-offs. On ViTacTip force regression and grating classification, TacEvo achieves high autonomous generation reliability (96.0%/94.5% trainable) and improves best validation fitness over 20 generations by 56.1%/96.1%. In a 20-seed post-search high-fidelity evaluation, TacEvo matches the expert baseline on force prediction and outperforms it on fine-grained grating classification. These results suggest that LLM-driven self-evolving search constitutes a practical paradigm for AI-assisted scientific discovery in specialised robotic sensing.
