Compare models

The main robot “brains” (VLA models) compared across the 7 capabilities of an autonomous robot. For each capability, we use the model's best published success rate on the same reference test. “Not measured” means no paper has published a result: it is not a zero.

Choose up to 3 models:

Capability by capability

For a like-for-like comparison, each capability is measured on the same reference test for every model. End of the gray bar = ideal robot (100%); vertical line = current record on this test. “≈”: reference test missing, average of the capability's other tests.

Basic skills

Can it carry out a simple instruction?

Test: Instruction following (LIBERO)

π0.5

98%

OpenVLA-OFT

97%

GR00T N1.5

97%

Robustness

Does it hold up when conditions change?

Test: Robustness to the unexpected (LIBERO-Plus)

π0.5

87%

OpenVLA-OFT

75%

GR00T N1.5

not measured

Realism

Would its results hold on a real robot?

Test: Small arm: realism (SimplerEnv · WidowX VM)

π0.5

73%

OpenVLA-OFT

42%

GR00T N1.5

68%

Two arms

Can it coordinate two hands?

Test: Two arms, cluttered scene (RoboTwin 2.0 · Hard)

π0.5

87%

OpenVLA-OFT

78%

GR00T N1.5

not measured

Home

Can it manage in an unfamiliar home?

Test: Virtual kitchen (mobile arm) (RoboCasa · Panda)

π0.5

62%

OpenVLA-OFT

44%

GR00T N1.5

66%

Memory

Does it remember what it has seen or done?

Test: Working memory (RoboMME)

π0.5

18%

OpenVLA-OFT

≈ 28%

GR00T N1.5

not measured

Long missions

Can it carry out a long mission on its own?

Test: Long, unpredictable missions (RoboCerebra)

π0.5

not measured

OpenVLA-OFT

not measured

GR00T N1.5

not measured

π0.5

Physical Intelligence · 2025 · ≈3 billion parameters

Open-source

The most versatile of the open models: designed to work in homes it has never seen.

Autonomy level: 1 / 5 Learns a skill
Capabilities measured: 6 / 7

On real robots (RoboArena), wins about 32% of its head-to-head matchups against the top-ranked model on the official leaderboard.

Test breakdown (12)

Instruction following98%
Two arms, cluttered scene87%
Two arms, clean scene87%
Robustness to the unexpected87%
Small arm: realism73%
Google robot: realism73%
Google robot: varied scenery68%
Common sense and general knowledge62%
Virtual kitchen (mobile arm)62%
Traps and perturbations62%
Virtual kitchen (humanoid)37%
Working memory18%

OpenVLA-OFT

Stanford · 2025 · 7 billion parameters

Open-source

OpenVLA retrained with a better recipe: much more accurate and about 25 times faster.

Autonomy level: 1 / 5 Learns a skill
Capabilities measured: 6 / 7

Test breakdown (10)

Instruction following97%
Two arms, clean scene80%
Two arms, cluttered scene78%
Robustness to the unexpected75%
Google robot: realism63%
Google robot: varied scenery54%
Virtual kitchen (mobile arm)44%
Small arm: realism42%
Short-term memory28%
Grasping varied objects19%

GR00T N1.5

NVIDIA · 2025

Open-source

NVIDIA's model designed for humanoid robots, trained at scale in simulation.

Autonomy level: 1 / 5 Learns a skill
Capabilities measured: 3 / 7

Test breakdown (6)

Instruction following97%
Small arm: realism68%
Virtual kitchen (mobile arm)66%
Google robot: realism52%
Google robot: varied scenery52%
Virtual kitchen (humanoid)48%