MambaAct Replaces Transformer Experts in Flow-Matching VLAs
October 15, 2026
MambaAct improves inference efficiency for flow-matching Vision-Language-Action models by replacing the Transformer-based action expert with a compact Mamba-based mixer. This maintains the visual-language prefix while reducing the computational bottleneck during iterative denoising.
HOW THIS AFFECTS YOU
●
builderYou can achieve lower latency for real-time robotic control by switching to Mamba-based action experts.
●
researcherReplacing Transformer-based action mixers with Mamba-based architectures shows promise for efficient VLA scaling.