Leaderboard
Physical AI, measured on real and simulated robots.
9 models on 4 benchmarks. Each benchmark has its own tasks and its own score.
| Model |
DROID: Everyday tasks
DROID
286 tasks · 1594 runs
Reliability% of task done
|
Pick and Place
Pick and Place
4 objects · 594 runs
Reliability% items placed
|
RoboLab-120
RoboLab
120 tasks · 2520 runs
Reliability% subtask score
|
MolmoSpaces-320
MolmoSpaces
320 tasks · 1920 runs
Reliability% success
|
|---|---|---|---|---|
Robot teleoperated by HumanTeleop
Reference
|
- | 100.0 | - | - |
OpenAI GPT-6 Astra, paused worldGPT-6 Astra, paused
Reference
|
- | - | 62.5±6.1 | - |
NVIDIA Cosmos3 NanoCosmos3 Nano
|
69.2±5.1 | - | 40.2±7.2 | 30.3±4.8 |
Black Forest Labs FLUX 3 ActionFLUX 3 Action
|
62.9±5.1 | - | 53.0±7.1 | 42.5±5.0 |
Ai2 MolmoAct2MolmoAct2
|
47.0±4.9 | - | 26.0±6.1 | 28.8±5.0 |
Galaxea G0.5G0.5
|
39.1±5.0 | - | 25.7±5.9 | 19.4±4.2 |
|
NVIDIA GR00TGR00T
|
36.0±4.8 | 72.3±4.6 | 35.9±6.8 | 22.5±4.7 |
Physical Intelligence π₀.₅π₀.₅
|
- | 75.1±4.8 | 40.9±6.0 | 19.4±4.4 |
Action Chunking TransformerACT
|
- | 63.7±7.0 | - | - |
Hugging Face SmolVLASmolVLA
|
- | 32.0±7.9 | - | - |
|
OpenAI GPT-6 AstraGPT-6 Astra
|
- | - | 7.0±2.8 | - |
± is half the 95% range. Each benchmark's page explains its score.
Updated