Robotics and Machine Vision
A CAD file and a camera in the cell, turned into a robot program your team can review. We scope it, build it and prove each stage on your parts.
Five Layers From CAD to Program
Each layer is its own piece of engineering, and each one can be tested alone.
Read the CAD
STEP geometry is loaded through the Open CASCADE kernel, so the system knows the exact shape of the part before it ever sees one.
See the Cell
A depth camera (RGB-D) watches the work area. Color alone struggles with shiny bare metal, and depth gives the system a shape to match.
Locate the Part in 3D
Detection and 6-DoF pose estimation are treated as separate problems. Open-vocabulary detection finds the part, and pose comes from the CAD model without per-part training.
Plan the Job
The AI drafts a structured plan from a library of proven routines. A schema check rejects anything malformed before it goes further.
Write the Program
A deterministic compiler turns the plan into motion, I/O and sequence. A validator and a simulated collision check run, and the output is the robot vendor's own program file.
The Engineering Underneath
Open tools where they fit, owned hardware for training, and an edge computer in the cell.
Motion and simulation
- ROS 2 and Python
- Motion planning and collision checking
- Simulation before anything reaches the controller
Training data
- Synthetic images from domain randomization and simulation
- Real images held back for validation only
- Accepted, edited and rejected programs fed back as labels
- Labeling checked against a written standard and known answers
Deployment
- Models trained on owned hardware
- Compressed for an edge device in the cell, then re-measured
- Open-source labeling tools with automatic pre-labeling
Voice drafting
A programmer can draft and edit a job by voice using a close-talk microphone and a custom vocabulary. The system reads the job back before anything is saved.
Built in Measured Stages
No stage is called done until it hits a number agreed in advance.
Foundation
CAD loading, the camera and the development environment, set up and checked against your parts.
See and locate
Detection and 3D pose on real parts in your cell, measured against the acceptance number.
Plan and validate
Structured plans, schema checks and simulated collision checks on your routines.
Emit and prove
Program files in your vendor's format, loaded by the vendor's own software and reviewed by your team.
Widen
More parts, more routines and voice drafting, added once the core loop is proven.
We build the vision and programming software. Robot safety systems, guarding and risk assessment stay with your integrator and your safety team.
Robotics and Vision Questions
Do we need to photograph and label every part?
Not for locating a part in 3D. Pose is estimated from the CAD model itself, without training on each new part. Detection is fine-tuned only when a part or cell needs it, with training images from simulation and real images kept to check the result.
Which CAD files do you accept?
STEP files first, read through the Open CASCADE geometry kernel. IGES can follow once STEP is proven on your parts.
Does the AI write code straight to the robot?
No. The AI drafts a structured plan and a schema check rejects anything malformed. A deterministic compiler writes the program, a validator and a simulated collision check run next, and a person confirms the result.
Will it work with our robot brand?
We write the vendor's own readable program file and let the vendor's software compile and load it. How much we can automate depends on how open that vendor's format is, which we check before quoting.
How is the project priced?
It is a custom quote, scoped stage by stage. Each stage ends with a measured result you can check before the next one starts.
Bring Us a Part and a Cell
Custom quote. Send a STEP file and a description of the cell, and we will scope the first stage with you.
