MambaAct swaps pi0.5 transformer action expert for compact Mamba
Original title: MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models
The public model keeps the pi0.5 vision-language encoder, shrinks the action expert 10 times, and raises mean LIBERO success by 1.08 points.
Authors introduce MambaAct, a Mamba action expert for flow-matching vision-language-action policies. It keeps the pretrained pi0.5 visual-language prefix encoder and replaces the transformer action expert. A Context Compressor maps prefix tokens into a compact signal that gates each Mamba block. They report 64.8% lower LIBERO denoising latency and 58.8% lower real-world latency, with code and weights public.
Unconfirmed: Latency and success figures are author-reported and not independently verified here.
- The action expert is about 10 times smaller than the transformer it replaces.
- LIBERO denoising latency falls 64.8%; real-world deployment latency falls 58.8%.
- Mean LIBERO success rises 1.08 percentage points across three fine-tuning seeds.
- Authors also report gains on LIBERO-Plus and real-world benchmarks.
- Code and weights are published on GitHub.
From the original article
1 min read at huggingface.co
Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference cost remains a major bottleneck for real-time control. This limitation is especially pronounced in flow-matching VLAs such as π_{0. 5}, where the action expert is evaluated repeatedly during iterative denoising.
By Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.