densereward-1frame

densereward-1frame

🌐 Project page · 📄 Paper (arXiv:2607.13033)

This is the single-frame reward model from DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation. It is a vision-language reward model for robot manipulation. Given a single image of a manipulation scene, it outputs a single scalar in [0.000, 1.000] estimating task progress (0 = no progress / failed, 1 = task complete).

Variants:

• densereward-1frame: single-frame reward model — given one RGB frame + task text, outputs a scalar reward in [0.000, 1.000].
• densereward-3frame-thinking: 3-frame reward model with reasoning — given 3 chronological frames + task text, emits a <think> reasoning word then a scalar reward. Warm-started from the 1-frame checkpoint.

Intended use & scope

• Input: one RGB frame of a robot manipulation scene (+ the task text in the user turn).
• Output: exactly one float with three decimals, e.g. 0.374.
• Not a chat model. It is trained to emit only a reward value under the system prompt below. It has no other instruction-following guarantees.

Output contract

The model was trained with a fixed system prompt and a strict output format. Use the same system prompt at inference (system_prompt.txt in this directory). It instructs the model to emit exactly a single float in [0.000, 1.000] (three decimals).

The user turn should contain the scene image plus the task description, in the same format used during training (image + short task text).

Quickstart

Requires transformers>=4.57, qwen_vl_utils>=0.0.14, torch, accelerate.

Tested environment: Python 3.12, CUDA 12.8, torch==2.8.0+cu128, torchvision==0.23.0+cu128, transformers==5.2.0, accelerate==1.13.0, qwen-vl-utils==0.0.14.

conda create -n densereward python=3.12 -y
conda activate densereward
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install "transformers==5.2.0" "accelerate==1.13.0" "qwen-vl-utils==0.0.14" pillow numpy

import re
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from qwen_vl_utils import process_vision_info

MODEL_DIR = "densereward/densereward-1frame"  # or a local path to this checkpoint

# The exact system prompt used in training ships with the model:
SYSTEM_PROMPT = open(f"{MODEL_DIR}/system_prompt.txt").read().strip()

model = AutoModelForImageTextToText.from_pretrained(
    MODEL_DIR, torch_dtype=torch.bfloat16, device_map="auto"
)
processor = AutoProcessor.from_pretrained(MODEL_DIR)

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "file:///path/to/frame.png"},
            {"type": "text", "text": "Task: put the black bowl on the plate."},
        ],
    },
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=8, do_sample=False)  # greedy

gen = out[:, inputs.input_ids.shape[1]:]
raw = processor.batch_decode(gen, skip_special_tokens=True)[0].strip()
reward = float(re.search(r"[01](?:\.\d+)?", raw).group())
print(raw, "->", reward)

Notes:

• max_new_tokens only needs to be a handful of tokens (a float is short).
• Always guard the parse: clip to [0, 1] and handle non-float output.

Loading with ms-swift

The release includes a minimal args.json ({model_type: qwen3_vl, swift_version}) so ms-swift's PtEngine / TransformersEngine can auto-detect the model type for this local directory (its architecture otherwise matches several swift registry entries). Load it as the base model with no adapter:

from swift import TransformersEngine, RequestConfig, InferRequest
engine = TransformersEngine("densereward/densereward-1frame")  # no adapters

Pass the system prompt above as a {"role": "system", ...} message and the task as the user turn (<image>{task}), matching training.

Compatibility note: ms-swift 4.2.x targets transformers>=4.57,<5. Under a much newer transformers (e.g. 5.x) the swift template/prompt composition can misbehave even though weights load fine — prefer the transformers path above, or a swift-matched transformers version, for swift-based inference.

License

Apache License 2.0 (see LICENSE). The base model Qwen/Qwen3-VL-4B-Instruct is also Apache-2.0. This fine-tune was produced at the University of North Carolina at Chapel Hill.

Citation

@article{fang2026densereward,
    title={DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation},
    author={Fang, Yu and Dong, Wanxi and Liu, Jiaqi and Yang, Yue and Huo, Mingxiao and Mu, Yao and Yao, Huaxiu and Li, Li Erran and Szafir, Daniel and Ding, Mingyu},
    journal={arXiv preprint arXiv:2607.13033},
    year={2026}
}