Natura-1 benchmark results
Natura-1 is evaluated on dexterous hands. GPT-6-Astra and π0.5 tested on the same robotic platform produced no measurable score, so their results on the ARX AC One dual-arm gripper platform are shown alongside.
Overall score
Percent, grouped by the robotic platform each model ran on
Tested on dexterous hands, but no measurable score
AgenticNatura-1
BaseGPT-6-Astraπ0.5
Average execution time
Seconds on one linear scale, fastest first. Further left is better
43.3× faster than GPT-6-Astra, 4.6× faster than π0.5
19.6× faster than GPT-6-Astra, 2.1× faster than π0.5
Scores by dimension
Natura-1 Agentic has the highest score, or ties for it, in all 12 dimensions. Each block lists the six rows in the order of the key. “No result” means the model was tested on that hardware but produced no measurable score.
Dexterous control
Control contact, coordinate both hands, and use tools.
Pose control & Contact
Bimanual coordination
In-hand manipulation
Tool use
Instruction following
Preserve intent across changes in scenes, objects, and actions.
Goal-directed reasoning
Spatial & scene generalization
Tool & object generalization
New-action generalization
Compositional generalization
Long-horizon execution
Organize actions, retain progress, and retry after failure.