Modelpublic

sii-research/tau-0-vla-world-model

sourceHugging Faceapache-2.0updated 5d agoView on Hugging Face

0likes

Model Card

τ₀-VLA World Model

Generate a goal image from a head-camera observation and a robot subtask. This repository contains the robotics LoRA for Step1X-Edit-v1p2.

bash

hf download sii-research/tau-0-vla-world-model robotics-lora.safetensors \
  --local-dir weights/world_model

Download the Step1X-Edit-v1p2 base model separately. Follow the World model guide for base-weight conversion, single-image and batch inference, and continuing LoRA fine-tuning. The adapter uses the original-format robotics implementation provided by that package.

The adapter has rank 64 and alpha 32. Default inference uses 28 denoising steps, CFG 6, size level 512, and seed 42. File checksums are in weights_manifest.json.

Use the Proposal model to predict subtasks from three camera views and task memory.

Code · Project · Paper