AlignIt

Model-Free Real-Time Robot Arm Alignment

Precisely align robot grippers with objects using RGB(D) cameras and neural networks. No CAD models, no markers, no complex setup—just record, train, and align.

Real-Time
Inference Speed
Model-Free
No CAD Required
RGB(D)
Camera Input

See AlignIt in Action

From data collection to autonomous alignment

Teaching Procedure

Record demonstration data as the robot moves around the target object. The system automatically collects and labels alignment data.

Continuous Alignment

Real-time inference predicts relative poses to precisely align the gripper with the object, enabling autonomous manipulation tasks.

How It Works

Three simple steps to autonomous alignment

1

Record an Object

Position your robot near the target object. The system automatically records camera images and corresponding poses as you demonstrate the alignment task.

2

Train the Model

AlignNet neural network learns the relationship between camera observations and required alignment actions using your demonstration data.

3

Align It

The trained model predicts real-time relative poses to guide the gripper to the target object with sub-millimeter precision.

Key Features

Designed for simplicity and precision

🎯 Model-Free Approach

No need for CAD models, 3D reconstructions, or object markers. AlignIt learns directly from camera observations and robot poses.

⚡ Real-Time Performance

Fast neural network inference enables smooth, responsive alignment even on standard computing hardware.

📷 RGB(D) Camera Support

Works with standard RGB cameras or RGB-D sensors like Intel RealSense for enhanced depth perception.

🤖 Robot Agnostic

Compatible with xArm, UFactory robots, and extensible to other manipulators through a simple robot interface.

🧠 Deep Learning Based

Powered by AlignNet, a convolutional neural network that predicts 6-DOF relative poses from camera images.

  • EfficientNet or ResNet backbones
  • Multi-view fusion support
  • Optional depth integration

🔧 Easy Data Collection

Automated spiral trajectory generation and data labeling makes dataset creation fast and effortless.

Technical Highlights

Built on proven technologies

Neural Network Architecture

AlignNet - A custom convolutional neural network that processes RGB(D) images to predict 6-DOF relative transformations.

  • Configurable CNN backbones (EfficientNet, ResNet)
  • Multi-view feature aggregation
  • Depth fusion for improved accuracy
  • 9D output (3D translation + 6D rotation)

Automatic Data Generation

Spiral trajectory generation creates diverse viewpoints around the target object for robust training.

  • Configurable cone angles and sweep ranges
  • Automatic pose labeling
  • Depth image recording
  • HuggingFace Datasets integration

Inference & Control

Real-time alignment loop continuously predicts and executes corrective motions until convergence.

  • Configurable tolerance thresholds
  • Rotation matrix acceleration
  • Convergence detection
  • Servo control integration

Applications

Where AlignIt excels

🏭 Manufacturing & Assembly

Precise part insertion, component assembly, and quality inspection tasks where sub-millimeter alignment is critical.

📦 Bin Picking

Align grippers with objects in bins or unstructured environments without needing explicit object models.

🔬 Laboratory Automation

Handle delicate samples, align with test fixtures, or perform precise liquid handling operations.

🎓 Research & Education

Teach vision-based manipulation, explore learning-based robotics, or develop custom alignment solutions.

Supported Platforms & Technologies

xArm / UFactory
Intel RealSense
PyTorch
HuggingFace Datasets
MuJoCo Simulation
Python 3.8+

Get Started with AlignIt

Open-source and ready to use

View on GitHub Contact for Support

Open-source project • Apache 2.0 License • Custom development available
Developed and maintained by Spes Robotics