Autonomous Waste Sorting for a Cleaner Future
GreenArm uses AI and robotics to detect, classify, and sort waste in real-time — reducing contamination and eliminating manual handling.
Waste mismanagement is a small systemic issue.
is incorrectly disposed, causing recycling contamination.
Manual sorting of hygienic waste poses serious health hazards.
Tissues and products flushed down drains clog entire systems.
Public and commercial washrooms generate enormous volumes of non-recyclable waste daily. Without intelligent sorting, contamination spreads through the entire stream — inflating costs, endangering workers, and failing our environment.
Four steps. Fully automated.
GreenArm orchestrates a seamless pipeline from detection to disposal — no human intervention required.
Detect
Overhead camera captures waste items using real-time computer vision.
Classify
YOLOv8 model identifies each object and assigns a category.
Pick
The Kinova robotic arm localises and grips the waste item precisely.
Sort
Item is deposited into the correct bin — Recycle or Compost.
Blueprint of an autonomous sorting cell.
Two defined zones. One intelligent system. Precise, repeatable, scalable.
Overhead Camera
180° waste detection field
ArUco Markers
Spatial calibration anchors
Waste Intake Area
Unstructured item placement
Kinova
Gen3 Arm
Recycle Bin
Paper, cardboard items
Compost Bin
Organic, tissue waste
Labelled Zones
ArUco marker-guided placement
ML, math, and robotics architecture from the Python implementation.
This section is sourced from the ROS2 Python stack in this repository (greenarm_perception, greenarm_manipulation, and kinova_gen3), not from placeholder web copy.
ML Pipeline (YOLO)
- Model: Ultralytics YOLO loaded from greenarm_perception/models/model_v9.pt.
- Inference runs per frame with configurable confidence threshold (default 0.5).
- Active classes are restricted to [0, 2], then the highest-confidence detection is selected.
- Bounding-box center is computed as ((x1 + x2) / 2, (y1 + y2) / 2) for target localization.
- Class-to-label mapping in the node: 0 → compost, 2 → recycle.
Math & Calibration
- ArUco-based homography calibration uses cv2.getPerspectiveTransform with four ordered marker points.
- Image-to-workspace transform is applied with cv2.perspectiveTransform to convert pixel centers into metric workspace coordinates.
- Marker pixel size estimate uses the mean of 4 side lengths via Euclidean norm.
- Workspace bounds enforce valid manipulation targets: x ∈ [0.2, 0.5], y ∈ [-0.3, 0.0] (meters).
- Legacy color pipeline applies HSV thresholding, morphology (open/close), contour area filtering, and normalized area confidence.
Control & Interfaces
- Perception publishes SourceTarget {x, y, z, confidence, label} to /source_zone/pick_target.
- Pick-place control uses a state machine: approach → descend → grasp → lift → move → drop → home.
- Target stabilization averages recent samples and accepts only low-deviation points (threshold 0.01 m).
- Drop poses are sampled uniformly inside class-specific destination zone bounds.
- Tool orientation is fixed for Cartesian moves at (theta_x, theta_y, theta_z) = (180, 0, 180).
Built for the hardest environments.
AI Vision Detection
YOLOv8 processes overhead camera frames in real-time, identifying waste type, position, and orientation with high accuracy.
Robotic Precision
The Kinova Gen3 arm executes sub-millimeter pick-and-place operations guided by ArUco marker calibration data.
Real-Time Sorting
End-to-end latency from detection to arm placement is minimal, enabling continuous, uninterrupted sorting cycles.
Hygienic Automation
Zero human contact with waste. Closed-loop automation eliminates exposure to biohazardous materials entirely.
Measurable outcomes. Real-world value.
GreenArm directly addresses the most costly and dangerous aspects of waste management.
Reduced Contamination
Precision sorting prevents cross-contamination between waste streams.
Safer Conditions
Workers are never exposed to hygienic waste or sharp objects.
Lower Maintenance
Proper disposal prevents plumbing failures and costly interventions.
Scalable Automation
Deploys across any facility size with minimal convenient configuration.
The roadmap ahead.
GreenArm is a foundation. The next iterations will push capability, accuracy, and autonomy further.
Expanded Object Classes
Train on a broader dataset to recognise more waste types — plastics, metals, glass.
Orientation-Aware Pickup
Smarter gripper planning that adapts to irregular shapes and awkward orientations.
Multi-Object Scheduling
Queue and prioritise multiple items for concurrent or sequential arm operations.
Improved Model Accuracy
Continual learning pipeline fed by real-world deployment data for iterative improvement.