Skip to content

fix(config): add PaliGemma weight loader to RoboArena configs#913

Open
wadeKeith wants to merge 1 commit intoPhysical-Intelligence:mainfrom
wadeKeith:fix/roboarena-weight-loader
Open

fix(config): add PaliGemma weight loader to RoboArena configs#913
wadeKeith wants to merge 1 commit intoPhysical-Intelligence:mainfrom
wadeKeith:fix/roboarena-weight-loader

Conversation

@wadeKeith
Copy link
Copy Markdown
Contributor

Problem

Fixes #856.

All five TrainConfig entries in roboarena_config.py are missing the weight_loader parameter. This causes them to default to NoOpWeightLoader, which initializes models from scratch instead of loading pretrained PaliGemma weights.

Fix

Added CheckpointWeightLoader to each RoboArena config:

  • Pi0FAST variants (binning, FAST, FAST specialist, FSQ): load from gs://openpi-assets/checkpoints/pi0_fast_base/params
  • Pi0 diffusion variant: load from gs://openpi-assets/checkpoints/pi0_base/params

Also added the missing import openpi.training.weight_loaders as weight_loaders.

Changes

  • src/openpi/training/misc/roboarena_config.py: Added weight_loaders import and weight_loader= parameter to all 5 TrainConfig entries.

Fixes Physical-Intelligence#856. RoboArena configs were missing weight_loader, causing models to
initialize from scratch instead of loading pretrained PaliGemma weights.
Added CheckpointWeightLoader for pi0_fast_base (FAST/binning/FSQ variants)
and pi0_base (diffusion variant) to match other config patterns.
@wadeKeith wadeKeith requested a review from kvablack as a code owner March 29, 2026 09:31
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

RoboArena config weight initialization missing

1 participant