zelpi.

How robot models score

LIBERO

A simulated Franka arm on 40 tasks in four suites: spatial layout, objects, goals and long multi-step tasks. Success rate in %, fine-tuned on each suite.

VLA-Adapter-Pro99.699.698.296.498.5VLA-Adapter paper
OpenVLA-OFT97.698.497.994.597.1OpenVLA-OFT paper
GR00T N1.697.598.597.594.597.0NVIDIA docs
π0.598.898.298.092.496.9openpi
UniVLA96.596.895.692.095.2VLA-Adapter paper
π096.898.895.885.294.2OpenVLA-OFT paper
GR00T N194.497.693.090.693.9VLA-Adapter paper
NORA92.295.489.474.687.9VLA-Adapter paper
SmolVLA90.096.092.071.087.3SmolVLA paper
π0-FAST96.496.888.660.285.5OpenVLA-OFT paper
SpatialVLA88.289.978.655.578.1VLA-Adapter paper
OpenVLA84.788.479.253.776.5OpenVLA-OFT paper
Octo78.985.784.651.175.1OpenVLA-OFT paper

Best score in each column highlighted. Select a column to sort.

Reported, not re-run. Numbers come from papers and model docs. Setups differ in cameras, training data and number of trials.

Simulation is a starting point. High LIBERO scores are common now; real robots and new scenes are harder. Test on your task in the playground.