Spike-driven Vision-Language-Action Model
Spike-driven VLA introduces a novel framework for robotic manipulation using spiking neural networks, aiming for energy efficiency over traditional large models. It combines spiking visual and instruction encoders with a Multi-Winner Spike Fusion module and a Spike Action Chunking Transformer to enable direct end-to-end training. This approach achieves competitive performance on benchmarks like LIBERO while requiring significantly fewer parameters and lower inference energy.