A lightweight, PyTorch-based Inverse Kinematics solver designed for robotics research and applications.

• Fast performance with Gauss-Newton optimization
• Parallel computation for solving multiple IK problems simultaneously
• Visualization capabilities for debugging and demonstration
• Simple structure with minimal dependencies
• Flexible architecture adaptable to various robot configurations
• End-to-end differentiability supporting gradient-based learning

pip install -r requirements.txt

python ik_example.py

This runs a demonstration using the Franka robot model included in the repository. The result is like that:

1. Create a RobotModel instance with your URDF file:

robot = RobotModel(
    urdf_path='path/to/your/robot.urdf',
    last_link_name='end_effector',
    load_mesh=True,
    device='cpu'
)

2. Define target poses and run batch inverse kinematics:

# batch size for optimization
n_samples = 1000
# Position and orientation (as quaternion) targets
targets = torch.tensor([[x, y, z, qw, qx, qy, qz]]).expand(n_samples,-1)

# Solve with multiple initial configurations for robustness
solutions = run_ik(robot, targets, n_samples=n_samples)

3. For visualization support, ensure your mesh files and link names align properly in the URDF.

• Leverages PyTorch for GPU acceleration and automatic differentiation
• Quaternion-based orientation representation
• Supports batch processing for parallel computation
• Simple customization through the RobotModel class

Please feel free to let me know if there are issues in runing the code.