Coordinates and errors are in pixels in the rotated 512 x 512 Aria frame. The goal is supplied in the prompt, so Aria FDE is partly degenerate and should not be interpreted independently of the other metrics.

Usage

python

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "wjdghks950/promac-aria-stage1b-cot-trajectory"
processor = AutoProcessor.from_pretrained(model_id, fix_mistral_regex=True)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Use the Aria structured-CoT v4 prompt schema and answer_only assistant format used during training. Evaluation used greedy generation with up to 512 new tokens.

Scope

This repository contains the merged Qwen3-VL inference model and processor/tokenizer. Trajectory inference uses the standard Qwen generate() path. It intentionally excludes the ProMAC future-state predictor, auxiliary reasoning decoder, SIGReg module, optimizer/scheduler state, raw .pt checkpoint, training logs, and datasets. It is not a resumable ProMAC training checkpoint.

The Stage-1b partial checkpoint was merged onto Qwen/Qwen3-VL-4B-Instruct revision ebb281ec70b05090aa6165b016eac8ec08e71b17. See provenance.json and SHA256SUMS for exact lineage and integrity information.

License And Data Terms

The upstream Qwen model card declares Apache-2.0. No standalone license was established locally for the ProMAC fine-tuned weights, so this repository uses license: other. Access to these weights does not grant rights to the Aria training or evaluation data. Users are responsible for complying with Project Aria and all applicable source-data terms.

Limitations

This model was evaluated on the documented Aria structured-CoT v4 validation split and may not generalize to other camera geometry, prompt formats, rotations, environments, or trajectory schemas. Reported metrics are offline evaluation results, not safety guarantees.