Researchers introduced MambaAct, a smaller action expert for flow‑matching vision‑language‑action (VLA) models. It keeps the visual‑language encoder of the earlier π₀.₅ model but replaces its transformer‑based action part with a compact Mamba module. The change reduces the action expert size by ten‑fold and cuts denoising latency by 64.8% on the LIBERO simulation benchmark and 58.8% on real‑world tests. At the same time, average success on LIBERO improves by about 1 percentage point. The code and model weights are released publicly.
Why it matters
Roboticists can run VLA models faster on limited hardware, enabling near‑real‑time manipulation.
