WAVE-0 — RoboTwin 2.0

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 RoboTwin 2.0 inference checkpoint, including policy weights and saved preprocessing/postprocessing statistics. The paper reports 91.47% overall success on RoboTwin 2.0, averaged across 50 tasks in both clean and randomized settings.

Model details

Setting	Value
Base model	lerobot/pi05_base
Training dataset	lerobot/robotwin_unified
Policy type / mode	`wave0` / `inference`
Action chunk / executed actions	32 / 32
Future prediction horizon	31 frames
Visual observations	Head camera and two wrist cameras
State / action dimensions	14 / 14
State and action normalization	Quantiles

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-RoboTwin --local-dir /path/to/models/WAVE-0-RoboTwin

# Cache the tokenizer before offline evaluation; accept its model terms if required.
hf download google/paligemma-3b-pt-224 --include 'tokenizer*' 'special_tokens_map.json'

Evaluate

Set up the simulator and assets using the official LeRobot RoboTwin guide. The repository provides an evaluation dependency reference.

ROBOTWIN_ROOT=/path/to/RoboTwin \
POLICY_PATH=/path/to/models/WAVE-0-RoboTwin \
NUM_GPUS=8 \
WORKERS_PER_GPU=4 \
bash scripts/eval/robotwin/eval_robotwin_all.sh

By default, this evaluates 50 tasks in both demo_clean and demo_randomized, with 100 episodes per task/setting and 32 executed actions per prediction. It uses dynamic multi-GPU scheduling and saves task logs and CSV/JSON summaries. Adjust NUM_GPUS and WORKERS_PER_GPU to fit your hardware.

For training recipes, see the repository README. For LIBERO, use the separate WAVE-0-LIBERO 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},
}