Deploy and operate OpenClaw at scale on Kubernetes using the official Helm chart. Covers multi-robot fleet management, horizontal pod autoscaling for simulation workloads, secret management, observability with Prometheus and Grafana, and zero-downtime upgrades.
Add the Helm repository: helm repo add openclaw https://charts.openclaw.dev && helm repo update. Install with defaults: helm install openclaw openclaw/openclaw -n openclaw --create-namespace. Customise values: helm install openclaw openclaw/openclaw -n openclaw --create-namespace -f my-values.yaml. Verify: kubectl get pods -n openclaw. The API server pod should reach Running state within 60 seconds. Access the API: kubectl port-forward svc/openclaw-api 7400:7400 -n openclaw.
In your Helm values file set: simulator: autoscaling: enabled: true minReplicas: 1 maxReplicas: 20 targetCPUUtilizationPercentage: 70. Apply: helm upgrade openclaw openclaw/openclaw -n openclaw -f values.yaml. The HPA controller will scale simulation pods between 1 and 20 based on CPU load. For GPU-accelerated simulation, configure KEDA with a custom metric from OpenClaw's Prometheus endpoint instead of CPU-based scaling.
Use Kubernetes Secrets: kubectl create secret generic openclaw-ai-keys --from-literal=OPENAI_API_KEY=your_key -n openclaw. Reference in values.yaml: envFrom: - secretRef: name: openclaw-ai-keys. For GitOps-safe secret management, use External Secrets Operator with AWS Secrets Manager, Azure Key Vault, or GCP Secret Manager as the backend. Never commit API keys in values.yaml or in Git history.
Enable metrics in values.yaml: metrics: enabled: true serviceMonitor: enabled: true. Install kube-prometheus-stack: helm install monitoring prometheus-community/kube-prometheus-stack -n monitoring --create-namespace. The OpenClaw ServiceMonitor will be picked up automatically. Import the OpenClaw Grafana dashboard from openclaw.dev/grafana-dashboard.json. Key metrics: openclaw_command_duration_seconds_bucket (latency histogram), openclaw_joint_error_total (error counter), openclaw_simulator_fps (simulation performance).
Update the image tag in values.yaml: image: tag: "3.2.0". Run helm upgrade openclaw openclaw/openclaw -n openclaw -f values.yaml. The rolling update strategy terminates old pods only after new pods pass readiness probes. The preStop hook (terminationGracePeriodSeconds: 30) gives in-flight commands 30 seconds to complete before the pod exits. Monitor the rollout: kubectl rollout status deployment/openclaw-api -n openclaw. Rollback if needed: helm rollback openclaw -n openclaw.
Deploy one OpenClaw API server per robot namespace (or use namespace-per-team with shared simulation infrastructure). Use a Kubernetes operator (openclaw-operator, available from the Helm chart) to manage robot CRDs: apiVersion: openclaw.dev/v1 kind: Robot metadata: name: arm-001 spec: model: ur5e ip: 192.168.10.101. The operator handles connection pooling, health checks, and automatic reconnection. A single 3-node Kubernetes cluster handles 50 robot connections with OpenClaw API pods at roughly 200 MB RAM per robot namespace.
Recommended values.yaml resource settings: API server: requests: cpu: 250m, memory: 256Mi; limits: cpu: 1000m, memory: 512Mi. Simulator pod (CPU): requests: cpu: 2000m, memory: 2Gi; limits: cpu: 4000m, memory: 4Gi. Simulator pod (GPU): requests: nvidia.com/gpu: 1; limits: nvidia.com/gpu: 1 (also set CPU and memory as above). Hardware bridge (real-time sensitive): requests: cpu: 500m, memory: 128Mi with CPU pinning via cpuManager policy: static.
For internal team access: use kubectl port-forward or configure an Ingress: values.yaml: ingress: enabled: true className: nginx hosts: - host: openclaw.internal.example.com paths: - path: / pathType: Prefix. Add TLS: tls: - secretName: openclaw-tls hosts: - openclaw.internal.example.com. For robot hardware outside the cluster, use a NodePort or LoadBalancer service type specifically for the hardware bridge port (7401) to avoid routing physical robot traffic through the Ingress.
Integrate OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 2.0 Flash, and self-hosted Ollama LLaMA 3 with OpenClaw for natural-language robot instruction and AI-powered manipulation planning. Includes detailed per-call and monthly cost breakdowns for each AI provider.
Deploy OpenClaw on AWS EC2 and EKS, Azure Virtual Machines and AKS, and Google Cloud Compute Engine and GKE. Includes detailed cost breakdowns per cloud, GPU instance sizing for simulation, and architecture patterns for hybrid cloud-to-robot connectivity.
Deploy OpenClaw inside Docker containers for reproducible, portable robotics development. This guide covers single-container setups, multi-service docker-compose stacks, GPU passthrough for simulation, and production image hardening.