austinpatel/bp_pi05_libero_gen_spatial_combination_lora

Paper: What Enables In-Context Behavior Prompting for Manipulation?

Code: real-stanford/behavior_prompting

pi05_libero_gen_spatial_combination_lora

$\pi_{0.5}$ LoRA finetuned checkpoint for LIBERO-Gen Spatial Combination, released with Behavior Prompting.

This is a raw openpi (Orbax) checkpoint: the repo contains the contents of a single training step directory (params/, train_state/, assets/, _CHECKPOINT_METADATA).

Train config (openpi fork)	pi05_libero_gen_spatial_combination_lora
Experiment name	pi05_libero_gen_spatial_combination_lora_seed0_v1
Step	99999
Training dataset	austinpatel/libero_gen_spatial_combination_train_openpi
Base model	pi05_base

Usage

Download into the location expected by the liberogen branch of the openpi fork:

# in openpi fork
hf download austinpatel/pi05_libero_gen_spatial_combination_lora --repo-type=model \
  --local-dir checkpoints/pi05_libero_gen_spatial_combination_lora/pi05_libero_gen_spatial_combination_lora_seed0_v1/99999

Then serve and evaluate it following the instructions in docs/libero_openpi.md. train_state/ is only needed to resume training; add --exclude "train_state/*" to skip it for inference.

License

These LoRA weights are released under Apache-2.0, like openpi. They are fine-tuned from pi0.5, which is built on PaliGemma, so use is also subject to the Gemma Terms of Use.