Robotic Policy Adaptation via Weight-Space Meta-Learning

Robotic Policy Adaptation via Weight-Space Meta-Learning introduces WIZARD, a weight-space meta-learning framework for vision-language-action robotics that generates task-specific LoRA adapters from a language instruction and a short demonstration video in a single forward pass, without action labels, test-time optimization, or privileged goal images. Built around the frozen π0.5 policy, WIZARD meta-trains on expert LoRA updates paired with multimodal task embeddings and adds two robotics-specific design choices: explicit per-layer scale prediction and cosine alignment in weight space. On the LIBERO benchmark, the method lifts zero-shot success on the held-out LIBERO-Spatial suite from 0.19 for the MT-VLA baseline to 0.40, while also improving LIBERO-Goal from 0.14 to 0.22 and LIBERO-Object from 0.01 to 0.03; the paper reports up to about 2× gains on unseen dataset collections and up to about 14× on unseen tasks. On a real Franka Emika Panda with three Intel RealSense cameras, WIZARD improves average success from 0.22 to 0.41 across five manipulation tasks after the same real-domain grounding, and its generated adapters can also warm-start fine-tuning to reach 96% expert success in 70 steps instead of 90. The ablation study shows that pure MSE reconstruction fails completely, scale-aware supervision raises average success to 0.27, and adding cosine alignment further increases it to 0.33, underscoring that stable parameter scale is essential for robotic weight generation.