WACO-pi05-LIBERO

Checkpoint of WACO (world-anchored canonicalization) for the LIBERO-Plus camera track, from the paper Towards Pose-Robust Few-Shot Vision-Language-Action Models.

Project page · Code · Dataset

Wenhao Li, Zhongyi Han, Yuhao Wei, Liang Han, Yilong Yin, Liqiang Nie (Shandong University · Shenzhen Loop Area Institute · Harbin Institute of Technology, Shenzhen · Xidian University)

What this is

A few-shot fine-tuned VLA silently assumes that the camera and the robot base stay exactly where they were during data collection. WACO recovers both shifts from a single RGB-D frame, with the arm placed by forward kinematics as a metric calibration target and the static scene as the world anchor, canonicalises the image to the calibration viewpoint and the state, action and start pose to the calibration base frame, and fine-tunes the unmodified policy under the artefacts of its own estimator.

This repository holds the policy weights of the row “WACO (error-aware fine-tuning + runtime)” of the LIBERO-Plus table of the paper:

  • Architecture: π0.5 as implemented in LeRobot (type: pi05), unchanged: no token, head or parameter is added.
  • Initialisation: the public π0.5 LIBERO checkpoint lerobot/pi05_libero_finetuned.
  • Training data: the public LIBERO demonstrations lerobot/libero, used without modification (one camera pose, one base pose; no perturbed training data).
  • Error-aware fine-tuning: 10k steps, 300 warm-up steps, global batch 32, learning rate 2.5e-5 with cosine decay to 2.5e-6, all weights (vocabulary embedding frozen), bf16. The augmentation is applied inside the dataset with p_rt = p_s = 0.5: round-trip reprojection to a random virtual camera (up to 12 cm and 12° per axis) and back with a shift perturbed by the estimator's error model (8 mm and 0.6°), rendered like the runtime's image channel on LIBERO-Plus (inverse warp, unseen pixels from the empty-workspace reference), and structured state noise (planar base-shift error, 8 mm and 0.6°). Action labels are unchanged.

Results

LIBERO-Plus camera track (1,599 tasks, one episode each), success rate in %. Released: the benchmark's prompts, which append the variant's parameters; corrected: the task's own instruction. Evaluation seed 0 / mean over three evaluation seeds.

Method Perturbed data Released Corrected
π0.5 public checkpoint – 62.5 63.8
+ WACO runtime (no training) – 89.4 / 88.0 88.1 / 87.7
WACO (error-aware fine-tuning + runtime) — this checkpoint – 93.7 / 92.4 93.7 / 92.0

The numbers of this checkpoint are obtained with the WACO runtime, which estimates the camera shift at reset and canonicalises the third-person image before it reaches the policy.

Usage

The checkpoint is a standard LeRobot π0.5 policy. To reproduce the LIBERO-Plus numbers, run it with the WACO runtime from the code repository:

git clone https://github.com/peacelwh/WACO.git && cd WACO
pip install "lerobot[pi,libero]==0.6.1" opencv-python-headless
bash scripts/install_plugin.sh        # registers --env.type=libero_poseshift with LeRobot
bash scripts/download_assets.sh       # calibration bundles for the runtime

# LIBERO-Plus camera track (in the LIBERO-Plus environment, see scripts/setup_liberoplus_env.sh)
POLICY=Bluexxx/WACO-pi05-LIBERO TAG=lp_waco CANON=waco PROMPT=released SEED=0 \
  bash scripts/eval_liberoplus_camera.sh
from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("Bluexxx/WACO-pi05-LIBERO")

Inputs: third-person image observation.images.image and wrist image observation.images.image2 (256 × 256), the 8-D state observation.state (end-effector pose in the base frame and gripper opening) and the task instruction. Output: 7-D end-effector deltas (n_action_steps = 10).

Files

model.safetensors (bf16 weights), config.json, the pre- and post-processor files with the normalisation statistics, and train_config.json (the training configuration; local paths removed).

License

This checkpoint is a fine-tune of π0.5 and is released under the terms of its base model (lerobot/pi05_libero_finetuned, Gemma terms of use). The code is released under Apache-2.0.

Citation

@article{li2026waco,
  title   = {Towards Pose-Robust Few-Shot Vision-Language-Action Models},
  author  = {Li, Wenhao and Han, Zhongyi and Wei, Yuhao and Han, Liang and Yin, Yilong and Nie, Liqiang},
  journal = {arXiv preprint},
  year    = {2026}
}
Downloads last month
29
Safetensors
Model size
4B params
Tensor type
F32
·
BF16
·
Video Preview
loading

Model tree for Bluexxx/WACO-pi05-LIBERO

Finetuned
(9)
this model

Datasets used to train Bluexxx/WACO-pi05-LIBERO