ypz21/GALA_robocasa_gr1

checkpoint_robocasa_gr1

GALA inference checkpoint checkpoint_robocasa_gr1 for RoboCasa GR1 tabletop evaluation.

Code and environment setup: PuzhenYuan/GALA (currently a private repository; access is required).

hf auth login  # Use an account with access to this private model repository.
hf download ypz21/GALA_robocasa_gr1 --local-dir checkpoints/checkpoint_robocasa_gr1
hf download Qwen/Qwen2.5-VL-3B-Instruct --local-dir checkpoints/Qwen2.5-VL-3B-Instruct

Run the parallel ID evaluation command in the code repository's README. This checkpoint requires the GALA inference loader and the Qwen2.5-VL processor/backbone files; it is not a standalone Transformers pipeline.

The export contains four safetensors shards, their index, inference configuration, and GR1 normalization/modality metadata. Optimizer and scheduler states, random states, training records, the bridge-supervision prediction head, and other embodiments' metadata are excluded. Inference weights retain their original precision. The exported policy produces identical actions to the original checkpoint for the fixed-input numerical validation.

Evaluation uses absolute joint actions, 16-step action chunks, four denoising steps, and 24 ID tasks with 50 episodes per task. Results are stochastic. The model is intended for research in simulation; physical robot deployment has not been validated by this release.

Qwen2.5-VL and simulator assets remain subject to their upstream licenses. See the code repository for code attribution and notices.