Fully open MolmoAct 2 tops independent robotics benchmark LIBERO-MAX on dynamic robustness

DJiafei · x · 2026-09-30

AI2's fully open robotics model MolmoAct 2, after task-specific fine-tuning, takes the top spot for overall task success on the independent Reality Check benchmark.

The new LIBERO-MAX benchmark tests dynamic robustness: a single environment change is introduced after the robot has begun acting — object/receptacle relocation, camera shift, sensor corruption, lighting or visual theme switch, distractor burst, or obstacle insertion — 1,000 matched cases per event type. Each Dynamic rollout is matched to a Base control with the same task, seed, and executed prefix, isolating the effect of the change. It measures whether task success is preserved, not whether the policy internally detects or deliberately replans.

The author cites Nicolas Keller's evaluation showing MolmoAct 2 still generalizes gracefully under these mid-episode perturbations.

Original post →

More from Embodied

Embodied channel →