crusbro robot-agent for Intelligent Robots
ROBOT · LLM + SKILLS + TOOLS · SAFE ORCHESTRATION
The intelligent core for robots in the embodied-AI era. As embodied intelligence moves from proof of concept to industrialisation, competition is shifting from mechanical structure and single performance metrics to the completeness of the intelligence stack. Built on the crusbro-agent SDK, this product sits on top of OS + ROS2, providing a perception–decision–execution loop with continuous iteration—turning natural-language goals into executable, auditable and recoverable workflows.
Intelligent core, redefining what intelligence means on a robot
The crusbro robot-agent sits on top of OS + ROS2 as the intelligent core of next-generation robots. It intelligently orchestrates capabilities, actions, dialogue, behaviour and emotional expression—intelligence made observable.
It fuses cloud model capabilities with on-device real-time perception and mechanical execution into a sense → decide → express → act → recover loop, moving robots from "able to move" to "able to understand intent, collaborate proactively, and serve continuously".
Industrial-grade embodied AI is judged by the intelligence stack
The way robots are evaluated is changing: from mechanical structure and single performance metrics to the completeness of the intelligence stack, with generality and transferability replacing heavy customisation as the core criteria.
Why an agent, not a cloud remote control
The value of a robot product is not "what model sits in the backend" but "whether the device sees, thinks and moves well". As the on-device intelligent core, crusbro robot-agent fuses cloud LLMs, Skills, Tools and real-time capabilities to turn intent into observable, auditable, recoverable behaviour.
Intelligence is a closed loop across sense, think, express and act. crusbro robot-agent assembles that loop from cloud models and on-device real-time capabilities, so the robot truly shows intelligence.
Seven-layer architecture with clean boundaries
Application → crusbro robot-agent orchestration → OS + ROS2 → capability layer → HAL → driver layer → hardware. Business apps never touch drivers or call LLMs directly; the capability layer exposes testable semantic interfaces through C++/Python SDKs and ROS2 Topics, Services and Actions, while HAL isolates differences between chassis, cameras, LiDARs and arms.
business systems · UI
planning · memory · sandbox · Skills/Tools
Topic · Service · Action
chassis · camera · LiDAR · arm
sensors · actuators
One orchestration model for robot applications
- Interaction, Perception, Task, Safety
- Identity, DeviceMgr, Nav, Content, Comm
- Boot, OTA, LogAudit, Power
- Apps submit tasks and context; the Agent picks Skills and calls authorised Tools
From goal to controlled action
- Understand the goal, identity and scene context
- Plan steps and verify permission, risk and device state
- Call navigation, vision, speech and arm capabilities
- Route critical actions through confirmation, approval or human takeover
- Persist the plan, tool calls, results and replay evidence
From "able to move" to "understands, collaborates, keeps serving"
| Stage | What the intelligent core does | What users observe as intelligence |
|---|---|---|
| Sense and understand | Fuses multimodal input with identity, memory and scene context to form intent. | The robot recognises you, follows context, and never asks the same question twice. |
| Think and plan | Decomposes goals into steps, picks Skills, and validates permission and risk. | Tasks advance on their own, with the robot explaining its key choices. |
| Express and interact | Selects expression strategies for dialogue, face, tone and motion amplitude. | Communication has pacing; motion has restraint—not cold, not excessive. |
| Act and recover | Drives navigation, vision, manipulation and arms; compensates, rolls back or hands over on failure. | Failures get an explanation, the action retries, and the robot never gets stuck. |
One core, many robot forms
The same intelligent core transfers across inspection, service, companion and logistics robots, with task experience reused across scenarios.
Cloud-edge collaboration with offline fallback
- Cloud · LLMs, Skills, permission policy, device management, audit and business integration
- Device · Keeps intelligence core: navigation, identity, anomaly response, task execution and safety control
- Transport · MQTT for commands / events / state; HTTPS for media and configuration
- Offline · Sense, decide, express and act still work; events replay by priority once reconnected
Security, privacy and rollback
- Device identity · X.509/mTLS, JWT, Topic ACL and HMAC anti-replay
- Data privacy · Identity and on-site sensitive data stay local first, encrypted upload per policy
- High-risk actions · Least privilege, HITL and full audit
- OTA & rollback · Signed OTA + A/B partitions; auto rollback on failure; maps and real-time control don't rely on cloud
Core value for robot products
Engineering delivery and acceptance
The integration prioritises proven chassis, depth cameras, LiDARs, mic arrays and arms, with HAL/SDK forming swappable vendor boundaries; every capability ships with fake / loopback implementations so CI runs without real hardware. Targets include cold start under 30s, interactive response under 2s, key anomaly detection under 1s, LAN MQTT latency under 200ms, and acceptance covers HIL stability, OTA drills, navigation traces, offline compensation and high-risk action auditing.