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.
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
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.
Train the Model
AlignNet neural network learns the relationship between camera observations and required alignment actions using your demonstration data.
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
Get Started with AlignIt
Open-source and ready to use
Open-source project • Apache 2.0 License • Custom development available
Developed and maintained by Spes Robotics