LoongForge-Pi05-LIBERO

This directory provides a π₀.₅ LIBERO SFT checkpoint trained with LoongForge.

π₀.₅ is a vision-language-action model that conditions a flow-matching action expert on a PaliGemma-style vision-language backbone. This checkpoint is a supervised fine-tune on the LIBERO dataset; the robot state is discretized into the text prompt and the policy emits end-effector delta actions in chunks of 10 steps.

Model details

Property	Value
Benchmark	LIBERO
Training data	LeRobot LIBERO
Inputs	Task instruction, external-camera image, wrist-camera image, and robot state
Robot state	8 dimensions (EE position 3, axis-angle orientation 3, gripper joints 2), encoded as discrete tokens in the prompt
Actions	7 dimensions per step (translation delta 3, rotation delta 3, continuous gripper 1); chunks of 10 steps
Image preprocessing	2 images, resize_with_pad to 224 × 224, siglip normalization at evaluation
Normalization	Mean/std normalization using the included dataset_statistics.json
Weight precision	Mixed BF16 / FP32, 813 tensors, 4.14 B parameters
Training schedule	6,000 iterations, global batch 256 (16 × 2 × 8 GPUs), base LR 2.5e-5 → min LR 5e-6, 1,000 warmup iterations, seed 1000
Reported result	98.75 % mean success over five independent 400-episode LIBERO evaluations (std 0.40, 95 % CI [98.2, 99.2])

Repository contents

model.safetensors          # Model weights
config.json                # Descriptive model / preprocessing / training metadata
dataset_statistics.json    # State and action normalization statistics
resume_meta.json           # Step count, epoch and checkpoint format

config.json is descriptive metadata only. It is not a Transformers AutoModel configuration — LoongForge still loads this checkpoint through its own explicit evaluation configuration.

Usage

Use the π₀.₅ implementation in LoongForge to load this checkpoint and run LIBERO evaluation. The entry point is examples/embodied/pi05/eval in the source tree.

Acknowledgements

Built on LoongForge, and LIBERO.