OpenAdapt Grounding

Robust UI element localization for automation.

Turn flakey single-frame detections into stable, reliable element coordinates.

The Problem

Vision models like OmniParser miss elements randomly frame-to-frame ("flickering"). Template matching breaks with resolution/theme changes.

Left: Raw detections showing frame-to-frame flickering

The Solution

1. Temporal Smoothing: Aggregate detections across frames, keep only stable elements
2. Text Anchoring: Match elements by OCR text (resolution-independent)

Side-by-side: Raw flickering (left) vs Stabilized detection (right)

Results

Detection Stability

Resolution Robustness

Visual Output

Stable elements after temporal filtering

Quick Start

uv pip install openadapt-grounding

Build a Registry (Offline)

from openadapt_grounding import RegistryBuilder, Element

# Add detections from multiple frames
builder = RegistryBuilder()
builder.add_frame([
    Element(bounds=(0.3, 0.2, 0.2, 0.05), text="Login"),
    Element(bounds=(0.3, 0.3, 0.2, 0.05), text="Cancel"),
])
# ... add more frames

# Build registry (keeps elements in >50% of frames)
registry = builder.build(min_stability=0.5)
registry.save("elements.json")

Locate Elements (Runtime)

from openadapt_grounding import ElementLocator
from PIL import Image

locator = ElementLocator("elements.json")
screenshot = Image.open("current_screen.png")

result = locator.find("Login", screenshot)
if result.found:
    # Normalized coordinates (0-1)
    print(f"Found at ({result.x:.2f}, {result.y:.2f})")

    # Convert to pixels
    px, py = result.to_pixels(width=1920, height=1080)
    print(f"Click at ({px}, {py})")

Run Demo

uv run python -m openadapt_grounding.demo

Output:

============================================================
OpenAdapt Grounding Demo Results
============================================================

Registry: 5 stable elements

📊 Detection Stability:
  Raw (with 30% dropout):    70%
  Stabilized (filtered):     100%
  Improvement:               +30%

📐 Resolution Robustness:
  ✓ 1.0x (800x600): 5 elements
  ✓ 1.25x (1000x750): 5 elements
  ✓ 1.5x (1200x900): 5 elements
  ✓ 2.0x (1600x1200): 5 elements

📁 Outputs saved to: demo_output/

How It Works

Temporal Clustering

Frame 1: [Login ✓] [Cancel ✓] [Password ✗]  → 2/3 detected
Frame 2: [Login ✓] [Cancel ✗] [Password ✓]  → 2/3 detected
Frame 3: [Login ✓] [Cancel ✓] [Password ✓]  → 3/3 detected
...
After 10 frames:
  - "Login" seen 9/10 times → KEEP (90% stability)
  - "Cancel" seen 7/10 times → KEEP (70% stability)
  - "Password" seen 8/10 times → KEEP (80% stability)

Text-Based Matching

At runtime, we use OCR to find text on screen, then match against the registry:

# Registry knows "Login" button exists
# OCR finds "Login" text at (0.45, 0.35)
# → Return those coordinates with high confidence

OmniParser Integration

Use with OmniParser for real UI element detection:

Deploy OmniParser Server

# Install deploy dependencies
uv pip install openadapt-grounding[deploy]

# Set AWS credentials (or use .env file)
cp .env.example .env
# Edit .env with your AWS credentials

# Deploy to EC2 (g6.xlarge with L4 GPU)
uv run python -m openadapt_grounding.deploy start

# Stop when done (terminates instance)
uv run python -m openadapt_grounding.deploy stop

Monitor Deployment

# Check instance and server status
$ uv run python -m openadapt_grounding.deploy status
Instance: i-0f57529053cb507ca | State: running | URL: http://98.92.234.13:8000
Auto-shutdown: Enabled (60 min timeout)

# Show container status
$ uv run python -m openadapt_grounding.deploy ps
CONTAINER ID   IMAGE               CREATED          STATUS          PORTS                    NAMES
c9343a65e85b   omniparser:latest   2 hours ago      Up 2 hours      0.0.0.0:8000->8000/tcp   omniparser-container

# View container logs
$ uv run python -m openadapt_grounding.deploy logs --lines=5
INFO:     99.230.67.57:61252 - "POST /parse/ HTTP/1.1" 200 OK
start parsing...
image size: (1200, 779)
len(filtered_boxes): 160 124
time: 4.438266754150391

# Test endpoint with synthetic image
$ uv run python -m openadapt_grounding.deploy test
Server is healthy!
Sending test image to server...
Found 5 elements:
  - [text] "Login" at ['0.08', '0.10', '0.38', '0.23']
  - [text] "Cancel" at ['0.08', '0.30', '0.38', '0.43']
  ...

Other Commands

uv run python -m openadapt_grounding.deploy build   # Rebuild Docker image
uv run python -m openadapt_grounding.deploy run     # Start container
uv run python -m openadapt_grounding.deploy ssh     # SSH into instance

Test Results

Real screenshot parsed by OmniParser:

Input	Output (160 elements detected)

Synthetic UI test:

Input	Output

# Run test with synthetic UI
uv run python -m openadapt_grounding.deploy test --save_output

Use OmniParser with Temporal Smoothing

from openadapt_grounding import OmniParserClient, collect_frames
from PIL import Image

# Connect to deployed server
client = OmniParserClient("http://<server-ip>:8000")

# Take a screenshot
screenshot = Image.open("screen.png")

# Run parser 10 times, keep elements in >50% of frames
registry = collect_frames(client, screenshot, num_frames=10, min_stability=0.5)
registry.save("stable_elements.json")

print(f"Found {len(registry)} stable elements")

Analyze Detection Stability

from openadapt_grounding import OmniParserClient, analyze_stability

client = OmniParserClient("http://<server-ip>:8000")
stats = analyze_stability(client, screenshot, num_frames=10)

print(f"Average stability: {stats['avg_stability']:.0%}")
for elem in stats['elements']:
    print(f"  {elem['text']}: {elem['stability']:.0%}")

UI-TARS Integration

UI-TARS 1.5 is ByteDance's SOTA UI grounding model (61.6% on ScreenSpot-Pro). Use it for direct element localization by instruction.

Deploy UI-TARS Server

# Install dependencies
uv pip install openadapt-grounding[deploy,uitars]

# Deploy to EC2 (g6.2xlarge with L4 GPU)
uv run python -m openadapt_grounding.deploy.uitars start

# Check status
uv run python -m openadapt_grounding.deploy.uitars status

# Test grounding
uv run python -m openadapt_grounding.deploy.uitars test

# Stop when done
uv run python -m openadapt_grounding.deploy.uitars stop

Use UI-TARS for Grounding

from openadapt_grounding import UITarsClient
from PIL import Image

# Connect to deployed server
client = UITarsClient("http://<server-ip>:8001/v1")

# Load screenshot
screenshot = Image.open("screen.png")

# Ground element by instruction
result = client.ground(screenshot, "Click on the Login button")

if result.found:
    # Normalized coordinates (0-1)
    print(f"Found at ({result.x:.2f}, {result.y:.2f})")

    # Convert to pixels
    px, py = result.to_pixels(width=1920, height=1080)
    print(f"Click at ({px}, {py})")

    # Optional: View model's reasoning
    if result.thought:
        print(f"Thought: {result.thought}")

OmniParser vs UI-TARS

Evaluation & Benchmarking

We provide a comprehensive evaluation framework to compare UI grounding methods.

Benchmark Results

Evaluated on synthetic dataset (100 samples, 1922 UI elements):

On Harder Synthetic Data (48 samples, 1035 elements - more dense, smaller targets):

Key Findings:

1. Cropping strategies dramatically improve UI-TARS accuracy (+96% with screenseeker) but have minimal effect on OmniParser on standard data
2. On harder data, cropping becomes essential - OmniParser baseline drops 7.3% but fixed cropping maintains 98%+ accuracy
3. OmniParser is 3.8-5x faster than UI-TARS while being significantly more accurate

Detection Rate by Method

Detection Rate by Element Size

Small elements (<32px) are hardest for UI-TARS (28.6% → 50% with cropping), while OmniParser maintains ~100% across all sizes.

Accuracy vs Latency Tradeoff

OmniParser offers the best accuracy-latency tradeoff, with near-perfect detection at <1.5s per element.

Synthetic Dataset Samples

The evaluation uses programmatically generated UI screenshots with ground truth:

Easy (3-8 elements)	Hard (20-50 elements, dark theme)

Run Your Own Evaluation

# Install dependencies
uv pip install openadapt-grounding[eval]

# Generate synthetic dataset
uv run python -m openadapt_grounding.eval generate --type synthetic --count 100

# Run evaluation (requires deployed servers)
uv run python -m openadapt_grounding.eval run --method omniparser --dataset synthetic
uv run python -m openadapt_grounding.eval run --method uitars --dataset synthetic

# With cropping strategies
uv run python -m openadapt_grounding.eval run --method omniparser-screenseeker --dataset synthetic
uv run python -m openadapt_grounding.eval run --method uitars-screenseeker --dataset synthetic

# Generate comparison charts
uv run python -m openadapt_grounding.eval compare --charts-dir evaluation/charts

Available Methods

See Evaluation Documentation for methodology and metrics.

API

RegistryBuilder

• add_frame(elements) - Add a frame's detections
• build(min_stability=0.5) - Build registry, filtering unstable elements

ElementLocator

• find(query, screenshot) - Find element by text
• find_by_uid(uid, screenshot) - Find element by registry UID

LocatorResult

• found: bool - Whether element was found
• x, y: float - Normalized coordinates (0-1)
• confidence: float - Match confidence
• to_pixels(w, h) - Convert to pixel coordinates

OmniParserClient

• is_available() - Check if server is running
• parse(image) - Parse screenshot, return elements
• parse_with_metadata(image) - Parse with latency info

UITarsClient

• is_available() - Check if server is running
• ground(image, instruction) - Find element by instruction, return GroundingResult

GroundingResult

• found: bool - Whether element was found
• x, y: float - Normalized coordinates (0-1)
• confidence: float - Match confidence
• thought: str - Model's reasoning (if include_thought=True)
• to_pixels(w, h) - Convert to pixel coordinates

collect_frames(parser, image, num_frames, min_stability)

• Run parser multiple times, build stable registry

analyze_stability(parser, image, num_frames)

• Report per-element detection stability

Documentation

Key Findings

From our benchmark on synthetic UI data:

• OmniParser dominates on our task: 97-99% detection vs UI-TARS's 36-70%
• Cropping becomes essential on harder data: OmniParser baseline drops to 90%, but fixed cropping maintains 98%+
• OmniParser is 3.8-5x faster than UI-TARS while being more accurate
• Literature benchmarks don't transfer directly: UI-TARS leads on ScreenSpot-Pro (complex instruction-following) but OmniParser wins on element detection
• Small elements (<32px) remain hardest for UI-TARS (28.6% baseline → 50% with cropping)

See Evaluation Findings for analysis of why results differ from literature benchmarks.

Development

git clone https://github.com/OpenAdaptAI/openadapt-grounding
cd openadapt-grounding
uv sync
uv run pytest

License

MIT