Any task. Any robot.
No training required.
the generalist policy for
physical intelligence
Iteration zeroIt works on day one.
It works on day one.
By day two, it’s 40× faster.
On any robot.
Humanoid
Mobile
Manipulator
0GPU-hours
task training
task training
Zeronce runs every benchmark zero-shot. The baselines all finetune to compete.
success rate % · 4 standard suites
| Model | Spatial | Object | Goal | Long | Avg | Training | Task training · GPU-h |
|---|---|---|---|---|---|---|---|
| Zeronce | Coming Soon | Zero-shot | 0 | ||||
| π0.5 | 97.0 | 99.0 | 98.0 | 96.0 | 97.5 | 6k-step finetune | — |
| OpenVLA | 84.7 | 88.4 | 79.2 | 53.7 | 76.5 | Finetune | — |
| Octo | 78.9 | 85.7 | 84.6 | 51.1 | 75.1 | Finetune | — |
Baselines as published. Zeronce evaluated zero-shot; our figures land here as evals complete.
Writings & research
From the zeroth iteration.
No writings yet
We’re running the benchmarks first. The write-ups land here once the numbers are in — check back soon.
00Coming soon
Papers in preparation
Protocols, ablations, and full per-suite results will be published here as our evaluations complete.