Qwen2-VL-2B-RobotArm, Leveraging Qwen/Qwen2-VL-2B-Instruct in Robotic Field

Model Description

This project focuses on fine-tuning visual language models to predict the position and orientation of a robotic arm for reaching a target object based on a human command and an image from a camera view. The system interprets natural language instructions and visual context to generate actionable pose outputs.

Usage

Inference with unsloth

from unsloth import FastVisionModel
from PIL import Image
from transformers import TextStreamer

model, processor = FastVisionModel.from_pretrained(
        'SakalYin\Qwen2-VL-2B-RobotArm',
        load_in_4bit=True,   
        local_files_only=True,
        attn_implementation="flash_attention_2",
    )

def resize_image(image_input, size=None):
    """Load image from path or matrices and resize them"""
    size = size if size else (854,480)
    if isinstance(image_input, str):  
        image = Image.open(image_input)
    else:  
        image = image_input

    size = size if size else (854,480)
    image = image.resize(size).convert('RGB') 
    return image

image_path = "https://i.imgur.com/vAleq1e.png"
size = (854, 480) # Recommeded Size
system_message = "You are a Visual Language Model Trained to output robot arm end-effectors parameters. Base on the user requests, locate the appropriate object in the image and you must return the position and orientation to reach it in xml format."
prompt = "Give me a cup."

image = resize_image(image_path, size=size)
messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": system_message if system_message else "You are a helpful assistant."}],
    },
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": prompt}]
    }
]

input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(
    image,
    input_text,
    add_special_tokens=False,
    return_tensors="pt",
).to("cuda")

# Generate text without using the streamer
text_streamer = TextStreamer(processor, skip_prompt=True)
generated_tokens = model.generate(
    **inputs,
    streamer=text_streamer,  
    max_new_tokens=128, 
    use_cache=True, 
    temperature=1.5, 
    min_p=0.1
)

# Decode output properly
output_text = processor.decode(generated_tokens[0], skip_special_tokens=True).strip()

Output: <obj>plastic_cup</obj> <pose>0.625,-0.785,0.548</pose> <orient>0.35478,0.71881,0.30505,0.5141</orient><|im_end|>

Results

Result compared to other model in the Collection

Model	Euclidean	X MSE	Y MSE	Z MSE	Roll Error	Pitch Error	Yaw Error
Qwen2VL-2B (Trained with Camera Location)	0.1089	0.0505	0.0575	0.0363	6.8334	5.4204	6.7619
Qwen2VL-2B	0.3865	0.1239	0.3411	0.0000	2.1462	0.9029	1.1926
Qwen2VL-7B	0.0509	0.0305	0.0320	0.0008	0.4148	0.1066	0.1734
LlaVA-NeXT 7B	0.0480	0.0306	0.0296	0.0000	0.0000	0.0000	0.0000
Llama Vision 11B	-	-	-	-	-	-	-