ROBOTIC POLICY ADAPTATION
via
WEIGHT-SPACE
META-LEARNING.

WIZARD reframes test-time adaptation as parameter inference: from one instruction and one short video, it generates a task-specific LoRA adapter in a single forward pass.

Christian Bianchi Siamak Yousefi Alessio Sampieri Andrea Roberti Luca Rigazio Fabio Galasso Luca Franco

ItalAI

Paper

Overview

TL;DR.

Standard VLA adaptation is bottlenecked by task-specific action labels and test-time optimization. WIZARD removes both: it predicts LoRA weights directly from language + video evidence, then injects them into a frozen policy for zero-shot execution.

The method is trained with a meta-dataset of pairs (task embedding, expert LoRA update) and adds three robotics-specific design principles: multimodal weight structuring, scale-aware generation, and alignment-oriented supervision in weight space.

WIZARD project teaser: Robotic policy adaptation via weight-space meta-learning

Method

How WIZARD Works.

BUILD TASK EXPERTS

Expert construction: train LoRA experts for meta-train datasets (LIBERO-Goal, LIBERO-Object, LIBERO-10) on top of a frozen VLA backbone.

LEARN WEIGHT MAPPING

Weight inference learning: encode task evidence from prompt and demo video, then train a meta-network to reconstruct expert LoRA updates in weight space.

ZERO-SHOT INFERENCE

Deployment-time adaptation: infer LoRA weights in one forward pass and inject them into the frozen policy, with no action labels and no gradient updates.

Multimodal structure LoRA tensors preserve vision / language / action boundaries.

Scale-aware synthesis Predict updates plus per-layer mean/std for stable control.

Alignment losses MSE + scale + cosine terms enforce functional directionality.

Evaluation

Zero-Shot Benchmark Results.

Experimental Protocol

Held-Out Shift

Split rule: meta-train on three LIBERO suites and evaluate on the unseen fourth suite.

Reproducible Evaluation

Fixed starts: each task is tested on 50 predefined initial states to preserve comparability.

Baselines

Reference methods: nearest-neighbor adapter retrieval and MT-VLA fine-tuning with OpenVLA-OFT and pi0.5.

Dataset	MT-VLA (pi0.5) Avg.	WIZARD Avg.	Delta
LIBERO-Spatial	0.19	0.40	+0.21 (~2.1x)
LIBERO-Goal	0.14	0.22	+0.08 (~1.6x)
LIBERO-Object	0.01	0.03	+0.02 (~3.0x)
LIBERO-10 (A/B subtasks)	0.03/0.03	0.09/0.07	+0.06/+0.04

Spatial reasoning| Bowl on stove to plate.

Object-centric transfer| Orange juice to basket.

Goal-conditioned behavior| Wine bottle to cabinet top.

Real Robot

Franka Emika Panda Transfer.

Both baseline and WIZARD start from the same pi0.5 checkpoint (+ 30 adaptation episodes), isolating gains from task-level weight generation.

Grasp and lift| Robust banana pickup from clutter.

Reach and pick| Stable cup pickup in real scene.

~2x

Gain on unseen dataset suites

~14x

Gain on unseen tasks (vs MT-VLA)

0.33

Real-world avg. success (vs 0.17)

Method	Banana	Apple	Marker	Cup	Apple→Cup	Avg.
pi0.5 baseline	0.27	0.13	0.10	0.30	0.07	0.17
WIZARD	0.53	0.33	0.17	0.63	0.17	0.33

Analysis

Data Efficiency & Warm Start.

Data Efficiency

On LIBERO-Spatial, WIZARD reaches 90% success with no task-specific gradient updates, while MT-VLA starts near 22% and needs 25 demonstrations to match that level.

Warm-Start Adaptation

Even where zero-shot fails, generated adapters provide a better starting region: expert-level performance is recovered in 7 steps vs 9 steps with MT-VLA initialization.

Citation

BibTeX.

@misc{bianchi2026roboticpolicyadaptationweightspace,
  title={Robotic Policy Adaptation via Weight-Space Meta-Learning},
  author={Christian Bianchi and Siamak Yousefi and Alessio Sampieri and Andrea Roberti and Luca Rigazio and Fabio Galasso and Luca Franco},
  year={2026},
  eprint={2606.07217},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2606.07217}
}

ROBOTIC POLICY ADAPTATION via WEIGHT-SPACEMETA-LEARNING.