Instructions to use InternRobotics/VLAC-Cut with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use InternRobotics/VLAC-Cut with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="InternRobotics/VLAC-Cut")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("InternRobotics/VLAC-Cut")
model = AutoModelForMultimodalLM.from_pretrained("InternRobotics/VLAC-Cut", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use InternRobotics/VLAC-Cut with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "InternRobotics/VLAC-Cut"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "InternRobotics/VLAC-Cut",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/InternRobotics/VLAC-Cut

SGLang

How to use InternRobotics/VLAC-Cut with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "InternRobotics/VLAC-Cut" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "InternRobotics/VLAC-Cut",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "InternRobotics/VLAC-Cut" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "InternRobotics/VLAC-Cut",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use InternRobotics/VLAC-Cut with Docker Model Runner:
```
docker model run hf.co/InternRobotics/VLAC-Cut
```

futurefantasy commited on 11 days ago

Commit

e801016

verified ·

1 Parent(s): 986931d

Refine VLAC-Cut model card style

Browse files

Files changed (1) hide show

README.md +32 -45

README.md CHANGED Viewed

@@ -14,37 +14,40 @@ license: other
 # VLAC-Cut
-**Video-based task progress estimation model for robot manipulation.**
 [Paper](https://arxiv.org/abs/2607.09776) · [Code](https://github.com/InternRobotics/VLAC-cut) · [Model](https://huggingface.co/InternRobotics/VLAC-Cut) · [Benchmark](https://huggingface.co/datasets/InternRobotics/VLAC-Cut-Benchmark)
-VLAC-Cut predicts how a robot task progresses over time from a natural-language instruction and an execution video. Instead of assuming that progress always increases with time, the model captures non-monotonic execution patterns including advancement, stagnation, regression, and recovery.
-The model is designed for long-horizon robot manipulation videos containing failed grasps, object drops, incorrect interactions, off-target placements, and corrective behavior.
 ## Highlights
-- **Video-level progress understanding:** reasons over robot videos rather than isolated images or image pairs, and cuts the video into good segments and bad segments.
-- **Non-monotonic process modeling:** recognizes advancement, stagnation, regression, and recovery.
-- **General understanding capability:** enables zero-shot generalization across tasks, scenes, and viewpoints.
-- **Coarse/Fine-grained flexible:** adjusts the frequency of video understanding, achieving progress analysis from coarse-grained to fine-grained.
-## Model at a Glance
 | Property | Description |
 |---|---|
 | Base model | `Qwen/Qwen3-VL-30B-A3B-Instruct` |
 | Input | Task instruction, optional task plan, and sampled video frames |
-| Output | Timestamped task progress |
-| Flexible video sampling rate | `2 Hz`-`20 Hz` (default: `2.0`) |
-## Model Design
-VLAC-Cut is fine-tuned from Qwen3-VL-30B-A3B-Instruct and retains its original multimodal generation interface. The model receives a task instruction, an optional task plan, and sampled video frames, then generates task progress at different timestamps as text.
-The training data also include supervision for robot behavior description, failure analysis, and correction planning, helping the model understand realistic robot execution processes.
-This repository contains the model checkpoint and loading assets only. Inference scripts, examples, and VPB evaluation code are maintained in the GitHub repository.
 ## Load with Transformers
@@ -78,25 +81,7 @@ python scripts/run_example.py \
   --output-jsonl <path-to-output.jsonl>
 ```
-Main arguments:
-- `--model-path`: Hugging Face model id or local checkpoint path.
-- `--video-path`: input video path.
-- `--task-instruction`: natural-language task description.
-- `--task-plan`: optional step-by-step task plan.
-- `--sample-hz`: video sampling rate. Default: `2.0`.
-- `--prompt`: optional full-prompt override.
-- `--output-jsonl`: optional output path; if omitted, the model response is printed only.
-Example response:
-```text
-时间: 0.0s, 进度: 0%
-时间: 1.0s, 进度: 25%
-时间: 2.0s, 进度: 40%
-```
-Render a prediction preview video:
 ```bash
 python scripts/utils/render_prediction_video.py \
@@ -106,18 +91,20 @@ python scripts/utils/render_prediction_video.py \
 ## Citation
 ```bibtex
 @misc{zhai2026maximizinghumanefficiencylargescale,
-      title={Maximizing Human Efficiency in Large-Scale Robot Post-Training via VLAC-Cut Guided Pipeline},
-      author={Shaopeng Zhai and Qi Zhang and Tianyi Zhang and Haoran Zhang and Fuxian Huang and Zhanhui Lin and Zijun Xu},
-      year={2026},
-      eprint={2607.09776},
-      archivePrefix={arXiv},
-      primaryClass={cs.RO},
-      url={https://arxiv.org/abs/2607.09776},
 }
 ```
 ## License
-See the license files distributed with this checkpoint. The base model and third-party training data remain subject to their original licenses and terms of use.

 # VLAC-Cut
+<p align="center">
+  <strong>Video-based Task Progress Estimation for Robot Manipulation</strong>
+</p>
+<div align="center">
 [Paper](https://arxiv.org/abs/2607.09776) · [Code](https://github.com/InternRobotics/VLAC-cut) · [Model](https://huggingface.co/InternRobotics/VLAC-Cut) · [Benchmark](https://huggingface.co/datasets/InternRobotics/VLAC-Cut-Benchmark)
+</div>
+## Overview
+**VLAC-Cut** is a video-language model for estimating the temporal progress of robot manipulation tasks. Given a natural-language task instruction, an optional task plan, and an execution video, the model produces timestamped progress estimates throughout the execution.
+Unlike methods that assume task progress increases monotonically over time, VLAC-Cut models non-monotonic execution dynamics, including advancement, stagnation, regression, and recovery. This formulation supports process-level analysis of partial completion, temporary failure, and subsequent recovery.
+This Hugging Face repository contains the VLAC-Cut model weights and loading assets. The official inference examples and evaluation code are maintained in the GitHub repository.
 ## Highlights
+* **Video-level temporal reasoning:** Analyzes robot execution videos rather than isolated images or image pairs and identifies temporal segments associated with task advancement or degradation.
+* **Non-monotonic progress estimation:** Captures advancement, stagnation, regression, and recovery without imposing a monotonically increasing progress assumption.
+* **Zero-shot generalization:** Generalizes across manipulation tasks, scenes, object configurations, and camera viewpoints.
+* **Flexible temporal resolution:** Supports configurable video sampling frequencies for both coarse- and fine-grained progress estimation.
+## Model Overview
 | Property | Description |
 |---|---|
 | Base model | `Qwen/Qwen3-VL-30B-A3B-Instruct` |
 | Input | Task instruction, optional task plan, and sampled video frames |
+| Output | Timestamped task-progress estimates |
+| Sampling rate | `2 Hz`-`20 Hz` |
+| Default sampling rate | `2.0 Hz` |
 ## Load with Transformers
   --output-jsonl <path-to-output.jsonl>
 ```
+Render a prediction JSONL file as an annotated video:
 ```bash
 python scripts/utils/render_prediction_video.py \
 ## Citation
+Please cite the following paper when using VLAC-Cut, the released model, or the Video Progress Benchmark:
 ```bibtex
 @misc{zhai2026maximizinghumanefficiencylargescale,
+  title         = {Maximizing Human Efficiency in Large-Scale Robot Post-Training via VLAC-Cut Guided Pipeline},
+  author        = {Shaopeng Zhai and Qi Zhang and Tianyi Zhang and Haoran Zhang and Fuxian Huang and Zhanhui Lin and Zijun Xu},
+  year          = {2026},
+  eprint        = {2607.09776},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.RO},
+  url           = {https://arxiv.org/abs/2607.09776}
 }
 ```
 ## License
+The model weights and third-party training data may be subject to additional licenses or terms of use. The source code in the GitHub repository is released under the MIT License.