Industrial Operations / Delivered project

Autonomous Robotic System for Nuclear Waste Sorting

A safety-led blueprint for identifying, grasping, and routing objects in controlled environments where human exposure should be minimized.

Explore the page ↓

Vision-guided robotic arm sorting industrial objects in a controlled laboratory — Industrial Operations · Data → Intelligence → ActionPhoto by Brecht Corbeel · Unsplash License ↗

01 / Problem statementDelivered project · Industrial Operations

The operating problem

Sorting unknown or hazardous materials is repetitive, variable, and potentially unsafe, while deterministic automation struggles with visual diversity.

Conditions the system must survive03 operating constraints

Objects may be unknown, occluded, reflective, or difficult to grasp

Unsafe actions must be prevented independently of model confidence

Remote operators need complete sensor and action evidence

02 / How we approached it

From operating uncertainty to testable evidence.

The work was decomposed into four engineering decisions. Each one produced an artifact the customer could inspect, test, and carry into deployment.

01Frame

Define safe object and action boundaries

Evidence produced

Object, action, and prohibited-state safety envelope

02Evaluate

Fuse vision, depth, and sensor evidence

Evidence produced

Vision and depth benchmark across difficult material conditions

03Engineer

Plan grasp and routing actions with confidence gates

Evidence produced

Grasp and routing validation in controlled scenarios

04Operationalize

Escalate uncertain items to remote human review

Evidence produced

Remote exception console with synchronized sensor and action trace

03 / Deployment record

Deployed around the workflow—not beside it.

The system boundary includes where inference runs, how evidence reaches existing tools, and how people handle uncertainty after launch.

Implementation statusDelivered project

Topology

Local robotic cell with edge perception and control, safety PLC boundaries, and a remote operator station.

Integration

Cameras, depth sensors, robot controllers, safety systems, inventory records, and exception workflows share synchronized state.

Operation

Confidence and safety gates stop uncertain actions; interventions, model versions, trajectories, and dispositions remain traceable.

04 / Evaluation metrics

What must be measured before the system earns trust.

Evaluation covers model behavior, workflow burden, and production performance. The metric defines the gate; the customer baseline and acceptance threshold define the target.

01Evaluation gate

Object perception recall

How it is measured

Detection and classification recall across material, occlusion, and pose scenarios.

What it decides

Defines when autonomous grasp planning is allowed.

02Evaluation gate

Grasp success rate

How it is measured

Successful secure grasps and placements by object class.

What it decides

Sets routing policy and retry limits.

03Evaluation gate

Unsafe action rate

How it is measured

Safety-envelope violations in simulation and controlled validation.

What it decides

Must remain at zero before any expanded operating scope.

04Evaluation gate

Human intervention rate

How it is measured

Remote interventions per item, with reason and cycle-time impact.

What it decides

Measures autonomy without hiding uncertainty.

05 / Customer perspective

Value has to appear in the customer’s operating day.

What matters in practice

The robot should stop or ask for help when evidence is insufficient; throughput is secondary to safe, traceable handling.

01Observable value signalReduced direct exposure to hazardous work

02Observable value signalMore consistent sorting evidence

03Observable value signalRemote handling of exceptions

04Observable value signalTraceable robot and model decisions

06 / Technology context

Tools follow the system—not the other way around.

Final architecture depends on data quality, operating conditions, integrations, risk, and evaluation criteria established during discovery.

RoboticsComputer visionDepth sensingEdge AISafety systems

Test the operating assumption

Define the evidence required to move from possibility to production.

Discuss this use case ↗

Autonomous Robotic System for Nuclear Waste Sorting

Sorting unknown or hazardous materials is repetitive, variable, and potentially unsafe, while deterministic automation struggles with visual diversity.

From operating uncertainty to testable evidence.

Define safe object and action boundaries

Fuse vision, depth, and sensor evidence

Plan grasp and routing actions with confidence gates

Escalate uncertain items to remote human review

Deployed around the workflow—not beside it.

What must be measured before the system earns trust.

Object perception recall

Grasp success rate

Unsafe action rate

Human intervention rate

Value has to appear in the customer’s operating day.

Tools follow the system—not the other way around.

Related operational systems.

Edge AI for predictive maintenance and leak detection

Computer vision for visual quality inspection

Define the evidence required to move from possibility to production.