Capstone: designing a bin-picking cell

This capstone pulls the course together in the design of a complete bin-picking cell, the most common application of the ideas in this course and one of the hardest to get right in production. You will work through the design the way an integration team does: start from requirements, choose an end effector, design perception, define the grasping and placing logic, size forces, budget cycle time and errors, plan for failures and complete the safety assessment. The worked numbers use a fictional but realistic scenario; the method applies to any cell.

The scenario: an e-commerce distribution center wants a robot cell that picks individual items from totes of mixed products and places them into outbound order boxes. The totes arrive on a conveyor, each holding 10 to 40 items of up to 30 different products. Items range from small boxed goods to polybagged clothing, bottles and blister packs, weighing 20 g to 1.8 kg. The client needs 700 picks per hour per cell on average, with at least 99 % of items handled without human help, and no damage to products.

Step 1: Requirements

Good cells start with written, measurable requirements, agreed with the client before hardware is ordered. Each requirement should say how it will be verified. For this cell, the key requirements are throughput (700 picks per hour sustained over a shift), automatic handling rate (at least 99 % of items without human intervention), damage rate (below 1 in 10,000 items), item range (the product mix above, with a list of excluded item types such as items over 1.8 kg or longer than 400 mm), availability (above 98 % of scheduled time), and safety (compliance with the applicable robot standards and workplace law).

Requirements also define what the cell does when it cannot pick an item: here, it moves the tote to an exception lane for a person, rather than trying indefinitely. Writing that down early prevents the most common integration argument, which is about what counts as success.

Step 2: End effector

Lesson 2's selection matrix applies directly. The product mix favors suction: most items have a flat or gently curved surface, and suction reaches items in the middle of a tote without fingers needing clearance. Polybags and blister packs seal well with soft bellows cups at moderate vacuum. Some bottles and irregular items are better held by fingers. The team chooses a combined head: a bellows suction cup on a short extendable stem, with two compliant fingers that can close around the item after suction lifts it, and a tool changer for a second, larger cup head used on heavy boxes.

Step 3: Perception

Following lesson 6, the cell uses an overhead active-stereo RGB-D camera about 0.9 m above the tote, which gives about 2 to 3 mm depth precision at the tote bottom. A learned instance segmentation model, fine-tuned on images of the client's products, separates items; a model-free grasp detector proposes suction points and finger grasps with scores. A plane fit around each suction candidate checks flatness, as in lesson 6. Transparent and highly reflective items, identified from the product catalog, trigger a second exposure with a polarizing filter and a more conservative grasp. The camera images the tote while the arm is placing the previous item, so perception overlaps with motion.

Step 4: Grasping and placing logic

The grasp logic ranks candidates by predicted success, height in the pile (topmost first, to avoid disturbing the pile), distance from tote walls, and a preference for item centers. Before execution, each candidate passes collision and reachability checks. The pick sequence is the one from lesson 8: pre-grasp 50 mm above the item, slow approach, suction on, lift 30 mm, check vacuum above threshold, close fingers if needed, retreat, move to the box, place, release with a blow-off pulse, verify with a quick image of the box. Each failure branch is specified: a vacuum failure triggers a retry on the next-best candidate, a second failure triggers re-imaging, and a third sends the tote to the exception lane.

# Throughput and automatic handling rate with retries, tote exchanges and exceptions (Monte Carlo)
import numpy as np
rng = np.random.default_rng(12)
items = 200000
p_first, p_retry, p_reimage = 0.93, 0.70, 0.60     # success on 1st try, retry and after re-imaging
t_cycle, t_fail, t_reimage = 3.0, 1.8, 1.2          # seconds: good pick, failed attempt, extra imaging
tote_items, t_tote = 25, 6.0                        # items per tote and seconds to exchange a tote
time = 0.0; auto = 0; exceptions = 0
for _ in range(items):
    if rng.random() < p_first:
        time += t_cycle; auto += 1
    elif rng.random() < p_retry:
        time += t_fail + t_cycle; auto += 1
    elif rng.random() < p_reimage:
        time += 2 * t_fail + t_reimage + t_cycle; auto += 1
    else:
        time += 3 * t_fail + t_reimage; exceptions += 1
time += (items / tote_items) * t_tote
print(f"automatic handling rate: {100 * auto / items:.2f} %")
print(f"exceptions per 1,000 items: {1000 * exceptions / items:.1f}")
print(f"sustained throughput: {3600 * auto / time:.0f} picks per hour")

automatic handling rate: 99.16 %
exceptions per 1,000 items: 8.4
sustained throughput: 1044 picks per hour

The simulation shows how the failure branches shape performance. A first-attempt success of 93 % alone would miss the 99 % automatic handling requirement, but one retry and one re-image lift it to about 99.2 %, and the sustained throughput stays well above 700 picks per hour even after tote exchanges. If first-attempt success fell to 88 %, the automatic rate would drop below 99 %; the requirement therefore translates into a concrete target for the perception and grasp team, which the acceptance test must confirm on the client's real product mix.

Step 5: Error budget

The fixed cell avoids mobile-base errors, but the error budget from lesson 10 still applies. Contributions are camera depth and lateral error at the tote bottom (about 2 to 3 mm), hand-eye calibration (about 1 mm after a careful calibration), robot repeatability (well under 0.1 mm for this class of arm, though absolute accuracy is worse) and the item shifting during approach. Bellows cups tolerate several millimeters of lateral error and about 10 to 15 degrees of tilt, so the budget is comfortable for suction; finger grasps on bottles need closer attention and benefit from the fingers' compliance.

Step 6: Failure modes and recovery

A failure modes and effects analysis lists what can go wrong, how it would be detected and what the cell does. Typical entries for this cell: double pick (two items lifted together), detected by weight on a wrist force-torque sensor or by the box camera, handled by returning one item; dropped item during transfer, detected by vacuum loss, handled by stopping and re-imaging the drop area; item placed outside the box, detected by the box camera, handled by re-picking; damaged packaging, detected by vision or by the vacuum signature of a torn bag, routed to the exception lane; camera blinded by glare, detected by depth image quality metrics, handled by retrying with a second exposure. Each entry has an owner and a test in the acceptance plan.

Step 7: Safety

In the United States, robot cells are governed by OSHA's general duty clause and machinery standards, with ANSI/A3 R15.06-2025, the US adoption of ISO 10218-1 and -2 (2025), as the recognized consensus standard. OSHA has no robot-specific standard, so compliance is judged against R15.06 and related standards such as ANSI B11.19 for safeguarding. The cell needs a documented risk assessment, safeguarding (here a perimeter fence with an interlocked access gate and a light curtain at the tote opening, monitored by a safety controller to the required performance level of ISO 13849-1), lockout procedures under 29 CFR 1910.147 for maintenance and training for operators and technicians.

If the client later wants people to work beside the robot without a fence, for example to clear exceptions, the risk assessment must be redone for collaborative operation, with speed and separation monitoring or power and force limiting, and the light curtain must stop the robot before a person can reach it. The safety distance formula from machine guarding (HSE-215) gives the minimum distance between the light curtain and the hazard.

Step 8: Acceptance testing

The design is proven by a site acceptance test, run on the client's real products under production conditions: for example, an 8-hour run with the full product mix, measuring throughput, automatic handling rate, damage and availability against the requirements in Step 1, with every exception logged and categorized. The test plan, written at the start of the project, defines the sample sizes needed for statistically meaningful results, as lesson 11 showed for learned policies. A cell that achieves about 99.2 % in simulation must demonstrate its rate over enough picks, typically several thousand, for the measured confidence interval to clear 99 %; the closer the true rate is to the requirement, the more picks are needed, which is a strong reason to design in margin.