KiyotakaWang/WAVE-0-LIBERO

WAVE-0 — LIBERO

                   

WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction

WAVE-0 (World-model-Augmented VLA Execution) brings lightweight future prediction to pretrained VLA policies. Built on π0.5, it uses a 257M-parameter latent world model to predict the visual consequences of latent action plans and condition action generation on those predictions. Its two-stage training learns action-grounded dynamics from robot demonstrations, then integrates the frozen world model into the policy, without pixel reconstruction or additional large-scale video pretraining.

This repository provides the LIBERO inference checkpoint, including policy weights and saved preprocessing/postprocessing statistics.

Results reported in the paper

Benchmark	Success rate
LIBERO	98.7%
LIBERO-Plus	88.0%
LIBERO-Pro	61.4%
LIBERO-Para	75.0%

The generalization benchmarks evaluate the LIBERO-trained policy without additional fine-tuning. See the paper for the individual protocols. The quick-start command below covers the original 40-task LIBERO benchmark, not the three generalization benchmarks.

Model details

Setting	Value
Base model	lerobot/pi05_base
Training dataset	HuggingFaceVLA/libero
Policy type / mode	wave0 / inference
Action chunk / executed actions	32 / 10
Future prediction horizon	9 frames
Visual observations	Two cameras, with one masked empty-camera slot
State / action dimensions	8 / 7
State and action normalization	Mean / standard deviation

Quick start

Use the custom WAVE-0 implementation in the code repository, built on LeRobot v0.4.4. Follow its installation instructions, run the following commands from the WAVE-0 repository root after installation.

Download

hf download KiyotakaWang/WAVE-0-LIBERO --local-dir /path/to/models/WAVE-0-LIBERO

# The evaluation script defaults to offline mode; cache this tokenizer first.
hf download google/paligemma-3b-pt-224 --include 'tokenizer*' 'special_tokens_map.json'

If tokenizer access requires accepting the model terms, complete that step on Hugging Face first. Download it using the same Hugging Face cache that will be used for evaluation.

Evaluate

Install the simulator and task assets following the official LIBERO guide into your WAVE-0 environment.

POLICY_PATH=/path/to/models/WAVE-0-LIBERO \
LIBERO_SRC=/path/to/LIBERO \
NUM_GPUS=8 \
WORKERS_PER_GPU=6 \
bash scripts/eval/libero/eval_libero.sh

By default, this evaluates all 40 tasks across the four original LIBERO suites, with 50 episodes per task, seed 42, and 10 executed actions per prediction. Adjust NUM_GPUS and WORKERS_PER_GPU to fit your hardware.

For training recipes, see the repository README. For RoboTwin 2.0, use the separate WAVE-0-RoboTwin checkpoint.

Citation

@misc{wang2026wave0bridgingvlapolicies,
      title={WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction},
      author={Haomin Wang and Anqiong Ji and Ke Wang and Zhongshan Liu and Ru Ying and Ran Zheng and Zhipeng Zhang},
      year={2026},
      eprint={2610.18568},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.18568},
}