AI Computer Vision Systems for Industrial Robots

Without vision, a robot is a blind automaton strictly following a program. CV adds adaptability: grasping arbitrarily oriented parts, in-process inspection, navigation in dynamic environments, and safe human collaboration per ISO 15066. We integrate computer vision systems into industrial manipulators and mobile robots — this combination delivers: 6DoF pose estimation for bin picking, depth-guided grasping, and semantic mapping for AMRs. Savings on a single bin picking station are substantial due to reduced manual labor and defects.

How CV Solves the Bin Picking Problem

Bin picking is one of the most demanded and challenging tasks. Parts in a container overlap each other, are chaotically oriented, and often have reflective surfaces. The primary method is 6DoF pose estimation: determining position (x,y,z) and rotation (roll,pitch,yaw) of each part. We use RGB-D cameras (RealSense, Azure Kinect) and one of the following models:

• FoundationPose — state-of-the-art for parts with a known CAD model. Achieves ADD-0.1d 78–89%.
• GDR-Net — geometrically discretized rendering, works without CAD but with lower accuracy.
• PVPN (Point Voting) — robust to heavy noise and partial occlusions.

Example implementation in PyTorch (abbreviated):

import numpy as np
import cv2
import torch
from dataclasses import dataclass
from typing import Optional

@dataclass
class ObjectPose:
    object_class: str
    position_xyz: tuple[float, float, float]    # mm in camera coordinate system
    rotation_matrix: np.ndarray                  # 3×3
    euler_angles: tuple[float, float, float]     # roll, pitch, yaw in degrees
    confidence: float
    grasp_point: tuple[float, float, float]      # recommended grasp point
    grasp_approach: np.ndarray                   # approach vector

class BinPickingSystem:
    """
    Bin picking system: detection and pose estimation of parts in a container.
    Methods:
    1. FoundationPose / DenseFusion: based on RGB-D
    2. GDR-Net: geometrically discretized rendering
    3. PVPN: point-wise voting
    Camera: Intel RealSense D435i or Azure Kinect.
    CAD model of the part is required for FoundationPose.
    """
    def __init__(self, pose_model_path: str,
                  cad_model_path: str,
                  object_classes: list[str],
                  camera_intrinsics: dict,
                  device: str = 'cuda'):
        self.device = device
        self.object_classes = object_classes
        self.camera_intrinsics = camera_intrinsics  # fx, fy, cx, cy

        # Load pose estimation model
        self.pose_model = torch.load(pose_model_path,
                                      map_location=device).eval()

        # CAD model for rendering (used by FoundationPose)
        self.cad_model = self._load_cad_model(cad_model_path)

        # YOLO for initial object detection
        from ultralytics import YOLO
        self.detector = YOLO(pose_model_path.replace('pose', 'det'))

    def _load_cad_model(self, cad_path: str):
        """Load .ply or .obj CAD model"""
        try:
            import open3d as o3d
            return o3d.io.read_triangle_mesh(cad_path)
        except ImportError:
            return None

    def estimate_poses(self, rgb: np.ndarray,
                        depth: np.ndarray) -> list[ObjectPose]:
        """
        Estimate object poses.
        rgb: (H, W, 3) uint8
        depth: (H, W) float32 in millimeters
        """
        # 1. Detect objects for ROI
        detections = self.detector(rgb, conf=0.4, verbose=False)

        poses = []
        for box in detections[0].boxes:
            cls_id = int(box.cls.item())
            if cls_id >= len(self.object_classes):
                continue

            x1, y1, x2, y2 = map(int, box.xyxy[0])

            # Crop RGB and depth patches
            rgb_crop = rgb[y1:y2, x1:x2]
            depth_crop = depth[y1:y2, x1:x2]

            if rgb_crop.size == 0:
                continue

            # 2. Pose estimation on the patch
            pose = self._estimate_single_pose(
                rgb_crop, depth_crop, cls_id, (x1, y1)
            )
            if pose:
                poses.append(pose)

        # Sort by Z height (top parts first)
        poses.sort(key=lambda p: p.position_xyz[2])
        return poses

    @torch.no_grad()
    def _estimate_single_pose(self, rgb_crop: np.ndarray,
                               depth_crop: np.ndarray,
                               cls_id: int,
                               offset: tuple) -> Optional[ObjectPose]:
        """Pose estimation for a single object"""
        from torchvision import transforms
        transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406],
                                  [0.229, 0.224, 0.225])
        ])
        from PIL import Image
        pil = Image.fromarray(cv2.cvtColor(rgb_crop, cv2.COLOR_BGR2RGB))
        rgb_tensor = transform(pil).unsqueeze(0).to(self.device)
        depth_tensor = torch.from_numpy(depth_crop).unsqueeze(0).unsqueeze(0).float().to(self.device)

        # Concatenate RGB + depth
        depth_norm = depth_tensor / 1000.0  # mm → meters
        # Simplified model takes 4-channel input
        input_tensor = torch.cat([
            rgb_tensor,
            torch.nn.functional.interpolate(
                depth_norm, size=rgb_tensor.shape[-2:], mode='bilinear'
            )
        ], dim=1)

        output = self.pose_model(input_tensor)
        # output: (1, 6) — translation(3) + rotation_euler(3)
        if output is None or output.shape[-1] < 6:
            return None

        out_np = output.squeeze().cpu().numpy()
        tx, ty, tz = out_np[:3] * 1000  # meters → mm
        rx, ry, rz = np.degrees(out_np[3:6])

        # Rotation matrix from Euler angles
        R, _ = cv2.Rodrigues(np.array([np.radians(rx),
                                        np.radians(ry),
                                        np.radians(rz)]))

        # Grasp point: object center + offset upward along normal
        grasp_z = tz - 30  # 30mm above center
        grasp_point = (tx, ty, grasp_z)
        approach_vec = R @ np.array([0, 0, -1])  # approach direction

        conf = float(torch.sigmoid(
            self.pose_model.confidence_head(output) if hasattr(
                self.pose_model, 'confidence_head') else torch.tensor(0.0)
        ).item()) if hasattr(self.pose_model, 'confidence_head') else 0.8

        return ObjectPose(
            object_class=self.object_classes[cls_id],
            position_xyz=(round(tx, 1), round(ty, 1), round(tz, 1)),
            rotation_matrix=R,
            euler_angles=(round(rx, 1), round(ry, 1), round(rz, 1)),
            confidence=round(conf, 3),
            grasp_point=grasp_point,
            grasp_approach=approach_vec
        )

Why 6DoF Pose Estimation Is Critical for Collaborative Robots

Collaborative robots work in the same space as humans. An error in determining the part pose leads to collision or damage. For cobot applications per ISO 15066, grasp repeatability of ±1 mm and latency under 50 ms are required. Only 6DoF pose estimation provides the accuracy needed for safe approach and grasp.

Method comparison metrics:

FoundationPose outperforms GDR-Net by 1.2× in ADD when CAD is available, but without CAD GDR-Net wins due to not requiring a model. PVPN is more robust to occlusions but slower (15 FPS vs 30 FPS for FoundationPose).

How We Design a Computer Vision System for Robots

The process starts with an audit of your production: what operations are performed, what parts, what current issues. Then we select hardware (cameras, lighting, controllers) and develop the CV algorithm. Steps:

1. Data collection: capturing scenes at your facility, labeling poses (6DoF) using our tools.
2. Model selection: FoundationPose, GDR-Net, or custom Transformer depending on CAD availability and acceptable latency.
3. Training and validation: on synthetic and real data. We aim for ADD-0.1d > 85%.
4. Integration: into a ROS2 node for the manipulator or OPC-UA for PLC. We ensure real-time performance.
5. Testing: on the production line for two weeks. We record KPIs (grasp cycle time, success rate).

Vision for AMR Navigation

Mobile robots (AMR/AGV) use CV for obstacle detection, people detection, and map building. Typical architecture: YOLOv8 on RGB-D, depth segmentation, sector division for trajectory planning. Example code snippet:

class AMRNavigationVision:
    """
    Computer vision for autonomous mobile robots (AMR).
    Tasks: obstacle detection, semantic mapping, human recognition
    for cobot safety (ISO/TS 15066 protected/restricted speed zones).
    """
    def __init__(self, obstacle_model_path: str,
                  device: str = 'cuda'):
        from ultralytics import YOLO
        self.obstacle_model = YOLO(obstacle_model_path)
        self.device = device
        # Semantic map: {cell_id: label}
        self.semantic_map: dict = {}

    def process_navigation_frame(self, rgb: np.ndarray,
                                   depth: np.ndarray) -> dict:
        """
        Analyze a frame for AMR navigation.
        Returns: obstacles, nearest_human_dist_m, clear_path_sectors.
        """
        results = self.obstacle_model(rgb, conf=0.4, verbose=False)
        obstacles = []
        nearest_human_dist = float('inf')

        h, w = depth.shape[:2]
        sector_width = w // 5  # 5 sectors: LL/L/C/R/RR

        for box in results[0].boxes:
            x1, y1, x2, y2 = map(int, box.xyxy[0])
            cls_name = results[0].names[int(box.cls.item())]
            cx = (x1 + x2) // 2
            cy = (y1 + y2) // 2

            # Median depth in bbox
            depth_crop = depth[y1:y2, x1:x2]
            valid_depths = depth_crop[depth_crop > 0]
            dist_m = float(np.median(valid_depths)) / 1000.0 if len(valid_depths) > 0 else 0

            obstacles.append({
                'class': cls_name,
                'bbox': [x1, y1, x2, y2],
                'distance_m': round(dist_m, 2),
                'sector': min(cx // sector_width, 4)
            })

            if cls_name == 'person' and dist_m < nearest_human_dist:
                nearest_human_dist = dist_m

        # Determine clear sectors
        blocked_sectors = {o['sector'] for o in obstacles if o['distance_m'] < 1.5}
        clear_sectors = [s for s in range(5) if s not in blocked_sectors]

        # ISO/TS 15066: if person < 0.5m → STOP; 0.5–1.5m → reduced speed
        safety_mode = ('STOP' if nearest_human_dist < 0.5
                       else 'REDUCED_SPEED' if nearest_human_dist < 1.5
                       else 'NORMAL')

        return {
            'obstacles': obstacles,
            'nearest_human_m': round(nearest_human_dist, 2),
            'clear_sectors': clear_sectors,
            'safety_mode': safety_mode
        }

What Is Included in the CV Project Work

We provide the full cycle: requirements analysis, equipment selection, camera-robot calibration, model training on your data, integration with the controller (ROS2/OPC-UA), testing under production conditions. Deliverables:

• Pose estimation model (ONNX/TensorRT)
• Integration module for PLC
• Safety operation documentation
• Operator training
• 6-month warranty support

Timeline and Cost

Timeline — from 8 to 20 weeks depending on complexity. Cost is calculated individually after an audit of your production. We guarantee stage transparency and fix KPIs in the contract.

Our Experience and Guarantees

Years of experience in industrial CV, dozens of projects from bin picking to inspection. Certified engineers in PyTorch, ROS2, OpenCV. We use official libraries: OpenCV and ISO 15066. We guarantee model accuracy (ADD and recall are specified in the contract).

Get a consultation for your project — contact us to discuss the task and create a prototype. Request a preliminary audit of your production — it is free.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.