Loading Engine Blocks With One 3D Camera and No Fixture
Machining an engine block is heavy work. The casting is large, the weight is high, and the block reaches the loading station on a forklift instead of in a fixture. Position and attitude differ with every delivery, so a taught position works until the next pallet arrives, and on a line running several block variants it has to be re-taught at every changeover, which is exactly when capacity disappears.
Random placement is the first obstacle, not the weight
The instinct is to solve handling first with a bigger robot and a stronger gripper. But a strong robot aimed at a position that has already moved does not help. The real constraint is that the pick point moves with every delivery and that variety multiplies the setups. Measuring each block before the grasp replaces mechanical positioning with measurement and compensation, without adding a fixture anyone has to load by hand.
One arm-mounted camera, four jobs
The vision system travels with the arm, so the arm is already the actuator for multi-angle imaging and the same hardware serves more than picking: it images before the grasp and confirms after the part is placed. Three further tasks fold into one cell:
- identify which block model arrived and load its template;
- measure the pose and guide the grasp into the machine fixture;
- confirm the machine is empty, and recognise an empty incoming pallet.
Combining them is not only about cost. One camera and one calibration chain mean one set of relationships to maintain, and imaging data is shared between tasks instead of bridging separate sensor systems, each with its own interface and cycle overhead.
Template for the model, local features for the pose
The chain runs from point cloud capture through feature recognition and pose computation to robot guidance. The system segments the block from the cloud, matches a template library of block models, and the match returns both identity and a coarse pose. A second stage extracts stable local features for fine positioning and produces the final grasp pose. Three calibrations, camera, arm flange and robot base, make that pose executable in robot coordinates.
The two checks that prevent collisions and dead cycles
Confirming the machine is empty is an interlock as much as a scheduling input: loading into an occupied machine causes a collision, waiting at an empty one wastes the cycle. The empty-pallet check works the other way, telling scheduling to replenish once the last block is taken, so the robot makes no futile approach and forklift supply matches consumption. Both reuse the same imaging path, and both rely on camera protection designed for chips and coolant mist.
30 seconds per part, 24 hours a day
The single-part cycle is about 30 seconds and the system runs continuously, 7 days a week and 24 hours a day, subject to site acceptance results. Continuous operation changes what must be designed in: calibration drift over long runs, occasional failed captures, and retry logic that recovers. A cell that assumes every capture succeeds stops at the first bad frame at three in the morning.
Loading cell questions
How does 3D vision handle a block that arrived off-position?
It measures the actual pose from the point cloud and converts the offset into the robot frame, so the path is compensated. Forklift placement accuracy stops being a constraint.
Does a model changeover stop the line?
No re-teaching is needed. The recognised model selects its own template and grasp parameters, and the switch happens in software.
What if several machines need serving?
The structure extends to multi-machine scheduling: the arm carries one vision system and visits machines on request instead of duplicating sensors.
What must be verified before 24/7 operation is accepted?
Failed-capture recovery, calibration stability over long runs and imaging consistency under shop dust. All three belong in acceptance testing, not in a maintenance plan written later.