The high-level planner of τ₀-VLA. It combines a task instruction, head and wrist camera images, and task memory to propose the next robot subtask.
This repository contains the complete Qwen3.5-9B model weights, tokenizer, and
processor. full_qa returns scene reasoning, updated memory, and a subtask;
subtask_only returns the next subtask as text.
hf download sii-research/tau-0-vla-proposal --local-dir weights/proposal
Use the Proposal guide for installation, three-camera inference, HTTP serving, and fine-tuning. The model uses camera order head → left → right. Supply your observations and inspect whether the proposed action fits the scene and task.
The world-model adapter
generates goal images from an observation and subtask.
File checksums are recorded in weights_manifest.json.
The high-level planner of τ₀-VLA. It combines a task instruction, head and wrist camera images, and task memory to propose the next robot subtask.
This repository contains the complete Qwen3.5-9B model weights, tokenizer, and
processor. full_qa returns scene reasoning, updated memory, and a subtask;
subtask_only returns the next subtask as text.
hf download sii-research/tau-0-vla-proposal --local-dir weights/proposal
Use the Proposal guide for installation, three-camera inference, HTTP serving, and fine-tuning. The model uses camera order head → left → right. Supply your observations and inspect whether the proposed action fits the scene and task.
The world-model adapter
generates goal images from an observation and subtask.
File checksums are recorded in weights_manifest.json.