Model card

Made by
Physical Intelligence
Size
3B parameters
Published
22 Apr 2025
More about the model
Publication
arXiv 2504.16054
Weights licence
Apache 2.0, with the Gemma Terms of Use source
Actions per chunk
π₀.₅: 50 source; π₀.₅ DROID: 15 source
Smallest GPU
More than 8 GB of GPU memory, for example an RTX 4090 source
Inference time
About 76 ms per chunk on an RTX 4090 (bfloat16, 5 denoising steps) source
Architecture
PaliGemma 2B vision-language model and a 300M flow-matching action expert source
Training data
Base model: 10k+ hours of robot data. π₀.₅ DROID: fine-tuned from it on the DROID dataset source

Run with Positronic

π₀.₅ DROID

Start the server source

docker run --gpus all -p 8000:8000 positro/pi05-droid

Drive a DROID robot from a positronic checkout guide

uv run --extra hardware positronic-inference real \
  --policy=.remote \
  --policy.address.host=localhost --policy.address.port=8000 \
  --next_task.instruction="Put the marker on the plate"

Progress through an episode

Each band shows how many episodes of π₀.₅ were in that band at each second. On a task of one step, an episode is in the band of the furthest stage it reached. On a task of several steps, an episode is in the band its credit has reached, and each band starts at its stage's credit for one step. The dashed line shows the mean Reliability.

Episodes

Each episode of π₀.₅ on DROID: Everyday tasks, the newest first.