Event
MambaAct uses a compact Mamba action expert to reduce flow-matching VLA inference latency
1 reports · 1 sources
MambaAct proposes replacing the Transformer-based action expert in π0.5 with a compact Mamba action expert while retaining its pretrained vision-language prefix encoder, in order to reduce the inference latency of flow-matching VLA models.
Report timeline
- MambaAct: Reducing Inference Latency of Flow-Matching VLAs with a Compact Mamba Action Expert
MambaAct proposes replacing the Transformer-based action expert in π0.5 with a compact Mamba action expert while retaining its pretrained vision-language prefix encoder.
arXiv cs.RO ·