Model or no model?
Learned models handle variety but can surprise you. Programmed approaches are predictable but break when something changes. We make the call separately for each stage of a task, and we show you the evidence.
Four rules we follow
Start from a scripted baseline
Every learned approach has to beat a programmed one on the same scenes, scored the same way. If it does not, the programmed one ships.
Count successes on your arm
Success rates from different labs are not comparable. 18 successes in 20 trials only tells you the true rate is somewhere between 68% and 99%, so we run enough trials to decide.
Learn only what geometry gets wrong
The useful question is not whether to learn, but which quantity is worth learning. Often a small learned part inside a programmed system does best.
Keep a programmed safety envelope
Current vision-language-action models output positions. None of them controls force and none can refuse, so workspace and force limits stay in code.
The call, stage by stage
Every pick-and-place task passes through the same four stages. Pick a stage to see both approaches and how we choose.
Stage 1 of 4: Perception
Where is the part, and which way is it facing?
Our usual call: Often learned
How we choose
If parts, fixtures and lighting are controlled, classical vision is cheaper and easier to certify. With many part types, clutter, or shiny and transparent parts, learned perception usually wins. We prove it on a held-out test set before we commit.
Programmed approach
- Fixed camera poses and fiducial markers (AprilTag, ArUco)
- Color and depth thresholds
- Template matching against CAD models
Learned approach
- Detection and segmentation models (RF-DETR, YOLO, SAM 2)
- 6-DoF pose estimation for known and new objects
- Trained on synthetic images plus a small real set
What tips the decision
Your constraints decide more than the model leaderboard does.
| Factor | Points to programmed | Points to learned |
|---|---|---|
| Part variety | A few known parts with CAD models | Many part types, or new ones every week |
| Scene | Fixtured parts and controlled lighting | Clutter, changing light, shiny or clear parts |
| Data | No demonstrations or images available | Demonstrations or simulated data are possible |
| Compute on the robot | Controller or PLC only | A GPU or Jetson-class computer next to the cell |
| Timing | Hard real-time loops (2 ms at 500 Hz) | Decisions that can take 30 to 100 ms |
| Certification | Every behavior must be explainable | Behavior can be bounded by a safety envelope |
Hybrid designs we use
Most systems we build are neither fully learned nor fully programmed.
Learned perception, programmed motion
A fine-tuned detector finds the part; MoveIt 2 plans a collision-free path to it. The most common design we ship.
Learned proposals, geometric checks
A grasp model proposes where to grip in a cluttered bin; collision and reachability checks reject anything unsafe.
Programmed approach, learned last centimeters
A planner handles free-space moves; a learned policy handles the contact-rich insertion or placement at the end.