OpenCastor is a universal runtime engine designed to provide intelligence to any robotic system. It offers a tiered cognitive architecture that scales from reflexive hardware-level safety to high-level strategic reasoning. By integrating with 12+ AI providers and supporting local execution via Hugging Face, Ollama, and llama.cpp, OpenCastor enables developers to build robots with semantic perception, long-horizon planning, and real-time reactive capabilities without vendor lock-in.

The runtime acts as an intermediary layer between physical hardware and cognitive AI models. It standardizes communication through the RCAN protocol, handles peripheral device integration automatically, and enforces safety bounds using a multi-layered kernel that includes emergency stops, collision avoidance, and tamper-evident audit logs.

Some of the key features are:

• Tiered Cognitive Architecture: Features a three-layer system separating reflexive safety, fast real-time perception, and high-level strategic planning.
• Plug-and-Play Hardware: Auto-detects USB cameras, LiDAR, IMU sensors, and motor controllers upon startup without custom driver code.
• RCAN Compliance: Implements the open RCAN protocol to ensure standardized robot identity, accountability, and communication across fleet devices.
• Flexible AI Integration: Supports cloud providers like OpenAI, Anthropic, and Google Gemini alongside local execution on Apple Silicon, Raspberry Pi, and NVIDIA Jetson hardware.
• Semantic Memory: Uses local CLIP embeddings to allow robots to recognize scenes, detect novel situations, and retain relevant past experiences.
• Safety Kernel: Enforces EU AI Act compliance through features like command identity, human-in-the-loop authorization, and physical bounds enforcement.
• Cross-Platform Compatibility: Runs natively on ARM64 and x86 architectures, supporting Raspberry Pi, Jetson, Linux, and macOS environments.
• Hot-Swap Brains: Allows switching between local and cloud AI stacks via simple configuration files to optimize for latency or cost.

Users begin by installing the OpenCastor runtime via a single command, which then probes the device to suggest an optimal AI stack. Configuration is managed through a single RCAN-compliant YAML file. The CLI tool allows users to scan for hardware, manage robot fleets, and run pre-tested recipes from the community. Once active, the runtime manages the interaction loop, routing commands between the local fast brain and the cloud-based planner, while maintaining strict adherence to safety protocol 66 (P66).

Some common use cases include:

• Home Security Patrol: Implementing autonomous robots with vision-based detection to monitor for open windows, water leaks, or unauthorized items with real-time alerts.
• Agricultural Monitoring: Deploying rovers to inspect crops for signs of pests, disease, or dry spots across large farm areas.
• Educational Robotics: Using common STEM kits in classroom settings to provide interactive Q&A bots or autonomous navigation exercises for students.
• Inventory Management: Equipping warehouse robots with QR and label scanning capabilities to track stock levels and report empty slots in real time.
• Teleoperation and Manipulation: Controlling robotic arms for precision tasks using leader-follower teleoperation presets and high-speed telemetry.