Abstract
WIZARD is a weight-space meta-learning framework that generates task-specific LoRA parameters for frozen VLA policies using language instructions and demonstration videos, enabling efficient task adaptation without fine-tuning.
Vision-Language-Action (VLA) models are emerging as a promising paradigm for robotic manipulation, enabling general-purpose policies trained from large corpora of demonstrations and action labels. However, adapting these models to new tasks still typically requires task-specific demonstrations, action annotations, and additional fine-tuning, making deployment costly and difficult to scale. We propose WIZARD, a weight-space meta-learning framework that sidesteps task-specific fine-tuning by generating task-specific LoRA parameters for a frozen VLA policy. Given only a language instruction and a short demonstration video, WIZARD predicts the corresponding adaptation weights in a single forward pass, without target-task action labels or test-time optimization. During meta-training, WIZARD learns to map task evidence directly to expert LoRA updates, capturing relationships between tasks in weight space. Experiments on LIBERO show that WIZARD improves performance by up to ~2x on unseen dataset collections and up to ~14x on unseen tasks. On a Franka Emika Panda, WIZARD consistently improves over a real-domain adapted baseline, showing that generated adapters provide task-level specialization beyond simulation.
Community
Paper Title
Robotic Policy Adaptation via Weight-Space Meta-Learning
Short Summary (TL;DR)
The authors present WIZARD, a framework that enables zero-shot robotic policy adaptation for large Vision-Language-Action (VLA) models without any test-time fine-tuning, online optimization, or action labels. Instead of tuning via gradients at deployment, a meta-network predicts task-specific LoRA parameters in a single forward pass from a language prompt and a short demonstration video. On the LIBERO benchmark, WIZARD improves success rates by up to 2x on unseen datasets and up to 14x on unseen tasks.
Suggested Community Comment
Title: 🚀 Zero-Shot LoRA Parameter Generation for Large VLAs
This paper introduces a clever workaround to the expensive, action-labeled fine-tuning usually required to adapt Vision-Language-Action (VLA) models to new tasks.
Why it's interesting:
- No Test-Time Gradients: WIZARD bypasses deployment fine-tuning entirely. It maps multimodal task embeddings directly to specialized LoRA parameters in a single forward pass.
- Scale-Aware Architecture: To stabilize weight generation across heterogeneous VLA modules, it introduces instance-wise token normalization and explicitly predicts layer-wise statistics.
- Strong Zero-Shot Baselines: It hits an average success rate of 40% on LIBERO-Spatial (vs. 19% for standard multi-task VLAs) and successfully transfers to a physical 7-DoF Franka arm, nearly doubling real-world success rates from 0.22 to 0.41.
It's a highly scalable approach to parameter generation that avoids full-policy weight synthesis while delivering serious data efficiency.
Definitely worth a read!
Get this paper in your agent:
hf papers read 2606.07217 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper