Every robot fleet generates data. An AMR with LiDAR produces 300 MB of point cloud data per minute. A vision-guided picking arm generates 1-2 GB of image data per hour. A fleet of 50 robots operating three shifts creates terabytes daily. Where that data gets processed, on the robot, at the facility edge, or in the cloud, determines your fleet's latency profile, operating cost, privacy posture, and scalability ceiling.
The cloud-versus-edge debate in robotics is not theoretical. It directly impacts how fast your robots react, how much you spend on connectivity, and how resilient your operation is when the internet goes down.
Architecture Overview
| Characteristic | Cloud Computing | Edge Computing | Hybrid | |---------------|----------------|----------------|--------| | Processing Location | Remote data center | On-robot or on-premises server | Split by task criticality | | Latency | 50-200 ms (typical) | 1-10 ms (on-robot), 5-20 ms (local server) | Varies by task | | Bandwidth Cost | High (all data uploaded) | Low (processed locally) | Moderate (selective upload) | | Compute Scalability | Virtually unlimited | Limited by local hardware | Best of both | | Offline Capability | None | Full (on-robot), partial (local server) | Partial | | Data Privacy | Data leaves facility | Data stays on-premises | Sensitive data stays local | | AI Model Updates | Centralized, instant | Requires deployment pipeline | Centralized training, edge inference |
Most production robot fleets in 2026 use a hybrid architecture. The question is not cloud or edge, but which tasks belong where.
When Cloud Computing Fits
Fleet Analytics and Optimization
Cloud platforms excel at aggregating data from multiple facilities, running large-scale analytics, and optimizing fleet behavior across an entire network. Locus Robotics' cloud-based fleet management system collects operational data from thousands of AMRs across hundreds of facilities to improve path planning algorithms, predict maintenance needs, and benchmark facility performance.
This cross-fleet intelligence is impossible to replicate at the edge. No single facility has enough data to train the optimization models that benefit from seeing patterns across millions of picks, thousands of facility layouts, and hundreds of operational configurations.
AI Model Training
Training computer vision models, manipulation policies, and navigation algorithms requires GPU clusters that are economically impractical to deploy at every facility. Cloud-based training pipelines from AWS RoboMaker, Google Cloud Robotics, and NVIDIA Isaac Sim allow robotics teams to train models on large datasets, validate performance, and push updates to edge devices.
A typical workflow trains a new picking model on 500,000 labeled images in the cloud (4-8 hours on an 8xA100 cluster), validates it in simulation, then deploys the inference model to edge devices. The training cluster costs $50-$100 per training run in the cloud versus $200,000+ to purchase equivalent hardware for on-premises use.
Multi-Site Fleet Management
Organizations operating robot fleets across multiple warehouses or factories need a central control plane. Cloud-based fleet management platforms from Formant, Freedom Robotics (now part of ABB), and InOrbit provide unified dashboards, remote monitoring, over-the-air updates, and cross-site reporting. A fleet manager in Chicago can monitor AMR performance in facilities across Memphis, Los Angeles, and Frankfurt from a single interface.
When Edge Computing Is Essential
Safety-Critical Decisions
Obstacle avoidance, collision prevention, emergency stops, and human detection must execute in single-digit milliseconds. A 150 ms cloud round-trip is unacceptable when a robot moving at 2 m/s needs to stop before hitting a person 0.3 meters away. Every production robot processes safety-critical perception on-board, using local processors like NVIDIA Jetson Orin, Intel RealSense depth cameras, or dedicated safety PLCs.
This is non-negotiable. No responsible robotics deployment relies on cloud connectivity for safety functions.
Real-Time Motion Control
Robotic arm trajectory planning, bipedal balance control, and precise pick-and-place operations require control loops running at 500-1,000 Hz (1-2 ms cycle time). These computations must execute locally. Even the fastest cloud connections introduce jitter and latency that would cause visible motion artifacts, dropped objects, or instability in legged robots.
Industrial robot controllers from FANUC, ABB, and KUKA have always processed motion control locally, and this will not change regardless of cloud computing advances.
Intermittent Connectivity Environments
Construction sites, agricultural fields, mining operations, and some warehouse cold storage facilities have unreliable network connectivity. Robots operating in these environments must function fully autonomously when disconnected. Edge computing ensures the robot continues to navigate, manipulate, and execute tasks without any cloud dependency.
Boston Dynamics Spot, deployed in remote inspection scenarios, operates entirely on edge compute during missions, uploading data only when connectivity is available.
Bandwidth and Cost Implications
The economics of cloud versus edge shift dramatically with fleet size and data volume.
| Scenario | Data Generated | Cloud Upload Cost (monthly) | Edge Processing Cost (monthly) | |----------|---------------|----------------------------|-------------------------------| | 10 AMRs, basic telemetry | 50 GB/month | $5-$15 | $0 (on-robot) | | 10 AMRs, full sensor data | 2 TB/month | $100-$200 | $50 (local server) | | 50 AMRs, full sensor data | 10 TB/month | $500-$1,000 | $150 (local server cluster) | | 50 AMRs + 10 vision arms | 25 TB/month | $1,500-$3,000 | $300 (local server cluster) |
Cloud upload costs include data transfer fees ($0.05-$0.12 per GB for major providers), storage costs, and the compute cost of processing the data. Edge processing costs are primarily the amortized cost of on-premises servers. At scale, edge processing costs one-fifth to one-tenth of equivalent cloud processing for high-bandwidth sensor data.
The practical approach is to process high-bandwidth, high-frequency data (LiDAR, cameras, IMU) at the edge and send aggregated metrics, events, and compressed summaries to the cloud.
The Hybrid Architecture in Practice
A well-designed hybrid architecture partitions workloads by latency sensitivity, data volume, and computational requirements.
On-robot (sub-10 ms): Safety systems, motion control, obstacle avoidance, localization, basic perception. Hardware: NVIDIA Jetson Orin, Intel NUC, or dedicated robot controller.
Facility edge (10-50 ms): Fleet coordination, traffic management, task assignment, local analytics, model inference for complex perception tasks. Hardware: 1-3 GPU-equipped servers (NVIDIA A4000 or equivalent), running Kubernetes or similar orchestration.
Cloud (50-500 ms acceptable): Model training, cross-facility analytics, long-term data storage, remote monitoring dashboards, over-the-air update management. Platform: AWS, Azure, or GCP with robotics-specific services.
This three-tier model is the architecture that major AMR vendors including Locus Robotics, MiR, and OTTO Motors deploy in production. It balances responsiveness, cost, and intelligence.
Implementation Recommendations
Start with edge-first design. Ensure every robot can operate fully autonomously without cloud connectivity. Then layer cloud services on top for analytics, training, and multi-site management. This order of operations guarantees operational resilience.
Invest in a robust local network. A dedicated 5 GHz Wi-Fi 6E network or private 5G installation is the backbone of edge-to-robot communication. Budget $15,000-$50,000 for proper coverage in a 100,000 square-foot facility. Poor local networking is the most common failure mode in fleet deployments.
Use cloud for what it does best. Aggregation, training, and monitoring are cloud strengths. Do not try to replicate them at the edge. Conversely, do not send raw sensor streams to the cloud unless you have a specific, cost-justified reason.
Plan for data sovereignty. If your facilities span multiple countries, edge computing simplifies compliance with GDPR, China's PIPL, and other data localization requirements. Keeping sensor data on-premises eliminates cross-border data transfer concerns entirely.
The hybrid model is not a compromise. It is the architecture that leverages each computing tier for what it does best, delivering responsive robots, intelligent fleets, and manageable costs.