LoongForge-VLDA-LIBERO

This repository provides a Stage 2 VLDA checkpoint trained on LIBERO using LoongForge.

VLDA builds on the π₀.₅ vision-language-action model with latent world modeling and world-conditioned action prediction. It uses V-JEPA 2 visual representations and a two-stage training procedure: world-model learning followed by joint world-model and action-policy training.

Model details

Property	Value
Benchmark	LIBERO
Training data	Physical Intelligence LIBERO
Inputs	Task instruction, external-camera image, wrist-camera image, and robot state
Robot state	8 dimensions, encoded as discrete tokens in the prompt
Actions	7 dimensions per step; chunks of 10 steps
Image preprocessing	224 × 224 for the policy inputs
Weight precision	Mixed BF16 / FP32
Normalization	Quantile normalization using the included q01 / q99 statistics

Repository contents

model.safetensors       # Model weights
vlda_config.json        # Model and data configuration
assets/norm_stats.json  # State/action normalization statistics

Usage

Use the VLDA implementation in LoongForge to load this checkpoint and run LIBERO evaluation. Follow the VLDA setup and evaluation instructions under cookbook/loongforge-vlda in the source tree.

In cookbook/loongforge-vlda/eval/configs/libero/libero.yaml, set server.ckpt_path to the downloaded checkpoint directory and server.dataset_statistics_path to its assets/norm_stats.json. Supply the PaliGemma SentencePiece tokenizer through server.tokenizer_path, and configure the local Python, repository, and output paths. Then run from the LoongForge root:

CUDA_VISIBLE_DEVICES=0 bash cookbook/loongforge-vlda/eval/run_libero_eval.sh

Keep the weights, configuration, and normalization statistics together. The tokenizer and LIBERO simulator assets are provided separately. This release contains model weights and configuration; it does not include optimizer state for resuming the original training run.

Acknowledgements

Built on LoongForge, OpenPI, V-JEPA 2, and LIBERO.