Ai2 just released MolmoAct 2: a fully open VLA-based architecture that runs 37x faster than the previous version and exceeds the Pi-05 model's performance on real-world robotics benchmarks. Full breakdown of the architecture, the OpenFAST tokenizer, and the 720-hour bimanual dataset they released alongside the model.
Source: https://arxiv.org/abs/2605.02881
Project page: https://allenai.org/blog/molmoact2
Code: https://github.com/allenai/molmoact2
Timestamps:
00:00 Intro
00:51 The problem with current VLAs
01:41 The core idea: adaptive 3D reasoning
02:39 How it actually works (4 components)
04:16 The 3 numbers that matter
05:09 What this means for builders
05:38 My take
06:15 Outro
Subscribe for more AI breakdowns
X: @sebuzdugan
Medium: medium.com/@sebuzdugan
Download
0 formats
No download links available.
MolmoAct 2 Performance: Comparing New Robotics AI Benchmarks | NatokHD