LoongForge-Pi05-RoboTwin

This directory provides a π₀.₅ RoboTwin 2.0 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 RoboTwin 2.0 bimanual demonstrations. The 14-dimensional robot state is discretized into the text prompt, and the policy emits chunks of 30 actions in π-space: arm values are deltas relative to the chunk-anchor state while the two grippers are absolute.

Model details

Property	Value
Benchmark	RoboTwin 2.0
Training data	Lerobot RoboTwin
Inputs	Task instruction, head camera image, left-wrist image, right-wrist image, and robot state
Robot state	14 dimensions of absolute π-space proprioception, encoded as discrete tokens in the prompt
Actions	30 actions per step, 14 dimensions per action
Image preprocessing	3 images, resize_with_pad to 224 × 224, siglip normalization at evaluation
Normalization	q99 quantile normalization using the included dataset_statistics.json
Weight precision	Mixed BF16 / FP32, 813 tensors, 4.14 B parameters
Training schedule	190,000 configured iterations, global batch 128 (8 × 2 × 8 GPUs), base LR 5e-5 → min LR 5e-6 cosine over 190,000 iterations, 2,000 warmup iterations, seed 1000
Reported result	77.56 % success over 50 tasks × 100 episodes (5,000 episodes) on demo_clean

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

Usage

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

Acknowledgements

Built on LoongForge, and RoboTwin.