MambaAct cuts flow-matching VLA denoising latency 64.8%

A new arxiv preprint, MambaAct, reports that swapping the transformer-based action expert inside the π₀.₅ flow-matching VLA for a Mamba-based module 10× smaller cuts denoising latency by 64.8% on LIBERO and 58.8% in real-world deployment, while nudging mean LIBERO success up by 1.08 percentage points across three fine-tuning seeds.

The design keeps π₀.₅'s pretrained visual-language prefix encoder in place. Inside the new expert, "Mamba serves as the action mixer that models dependencies across the noisy continuous-action sequence," the authors write, and a lightweight Context Compressor maps VLM prefix tokens into a conditioning signal "that gates and modulates each Mamba action block" so cross-attention does not have to run on every denoising step.

The authors — Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari and Cesare Stefanini — frame the problem as inference cost: in flow-matching VLAs, "the action expert is evaluated repeatedly during iterative denoising," which is where the latency bill accumulates. Gains are also reported on LIBERO-Plus and on real-world benchmarks, though the abstract publishes no absolute millisecond figures and does not name the real-world hardware. Code and weights are posted on GitHub.