Helooo,
I’ve been working on hold_and_weld, an open-source ROS 2 framework for dual-arm robotic welding. The goal is to reduce reliance on dedicated welding fixtures by using one robot to grip and hold the workpiece while another welds it.
video:
https://github.com/user-attachments/assets/986d984e-00fb-4769-97a7-4cb6e9866428
The grasping and welding components can also be used independently.
The framework combines constraint-aware grasp sampling, weld seam extraction from CAD/mesh geometry, and motion planning with MoveIt 2.
It targets ROS 2 Jazzy on Ubuntu 24.04. It’s currently a proof of concept demonstrated in simulation. Repository
Constraint-aware grasp sampling
The grasp sampler generates antipodal grasp candidates on exact CAD surfaces. Exclusion volumes keep grasps away from weld seams, screw holes and other forbidden regions. Candidates are filtered using gripper geometry and checked for collisions with the workpiece, secondary objects and the ground. STEP and URDF inputs are supported.
Weld seam extraction
The weld finder extracts seams from CAD and mesh geometry and exports weld paths as JSON. The mesh pipeline identifies contact boundaries between parts, traces them along the parts’ edges, and refines the resulting weld paths.
Weld approach configuration and motion planning
Reaching the start of a seam does not necessarily give the robot a suitable configuration for completing the weld. ConfigurationFinder selects the approach joint configuration before OMPL plans the approach.
A custom Ceres-based IK solver minimizes position and orientation errors, with hard joint bounds and a seed penalty that favors configurations close to the seed. Multiple initial seeds explore alternative IK branches.
Each candidate is evaluated along a sampled Cartesian path covering the plunge and weld motion. IK is solved at each sample using the previous solution as the seed. Candidates are rejected on IK failure, joint-limit violations, low Yoshikawa manipulability or excessive joint steps. Feasible candidates are ranked by their minimum joint-limit margin, minimum manipulability and distance from the home configuration.
OMPL plans to the selected configuration as a joint goal. Pilz handles the plunge and linear/circular weld motions; other curves use a dense-waypoint Cartesian fallback.
Dual-arm coordination
Action servers handle grasping and welding, while a coordinator sequences the workcell: the gripper arm picks and holds the workpiece, then the welding arm executes the weld path. Separate launch files allow either side to run independently.
Robot and gripper catalog
Models and robot base poses are selected through workcell.yaml.
Additional robot and gripper models can be easily integrated through xacro wrappers and catalog registration. A guide documents the required interfaces and configuration steps.
Configurator and visualizers
The workcell configurator lets you select robot and gripper models and position robots and scene objects using interactive markers in RViz2. Layouts are saved as YAML files, which can be copied into the workcell configuration.
Separate visualizers display grasp candidates and weld paths.
Testing and current limitations
The repository includes unit tests. I also created a complex geometry example with concavities, elliptical surfaces and other geometric features to test grasp sampling and seam extraction. Results look good on the tested examples; additional geometry testing would help assess broader coverage.
The complete workflow has been demonstrated in simulation. It has not yet been validated through physical welding trials.
Existing physical demonstrations show that robot-held welding is feasible at smaller scales too. I’m actively working toward a hardware demonstration of this framework.
The roadmap outlines planned development.
Installation and documentation
Prebuilt Docker image:
docker pull ghcr.io/silanus23/hold_and_weld:0.3.0
The repository includes installation instructions, usage examples, package documentation and a guide for adding robot and gripper models.


