Head Intelligence
Edge AI Compute
Jetson Orin NX + FPGA
Vision · Audio · Gaze
Human interaction
Humanoid Robot
Architecture
A reference architecture that unifies perception, deterministic control, functional safety and intelligent interaction in one scalable system.
Body AI
NVIDIA Jetson AGX Thor
Task · Perception · Planning
Context-aware control
Real-Time Motion
Real-Time Controller
Balance · Gait · Joint control
EtherCAT Master
Safety Architecture
Independent Safety Controller
Controlled stop · Brake / hold
Fall-risk management
Head ↔ Body AI Perception · Speech · Gaze targets
Motion ↔ Drives Setpoints · Encoder / torque feedback
Safety ↔ Drives Stop category · Brake / hold · Status
System Architecture Views
Four views of sensing, timing, compute, motion and safety across the humanoid platform.
01
Whole-Body Architecture
A synchronized pipeline connects sensing and system telemetry to accelerated AI, deterministic motion and independent safety control.
Recommended Albatron platforms
AI perception and planning.
Synchronized sensor acquisition.
Time-aware network connectivity.
EtherCAT inertial-state feedback.
Motion-control host; requires a real-time OS, control stack and cycle validation.
Safety scope: This is a reference safety concept, not a certified safety function. The safe state and stop strategy—including controlled stops, brake / hold behavior, drive isolation and fall-risk management—must follow a project hazard analysis and be validated to applicable standards such as IEC 61508, ISO 13849 and IEC 61800-5-2.
02
Sensing, Timing & Motion Architecture
Sensor time, AI planning and deterministic joint execution remain separate, then meet through an explicit PTP-to-EtherCAT clock correlation.
Time-aligned sensor dataTrajectory / setpointsEtherCAT DC control loopIndependent safetyPower distributionDashed = return / feedback
Cameras · Microphones
Native IMU · Neck encoders
Cameras · LiDAR
Native IMU · System telemetry
Palm vision · Tactile
Force · Joint encoders
Native IMU · Force / torque
Joint encoders
Synchronized data plane
IEEE 1588 PTP or IEEE 802.1AS gPTP · Deployment-specific profile
Time referenceDisciplined network clockEdge timestampsAssigned near sensor captureData transportTimestamped sensor streams
NVIDIA Jetson AGX Thor
Perception · Planning · Behavior
State estimation · Interpolation
Cyclic control · EtherCAT master
4.EtherCAT DC
cyclic setpointsEncoder / torque / EtherCAT IMU feedback
cyclic setpointsEncoder / torque / EtherCAT IMU feedback
Position · Velocity · Torque loops
Encoders · Drive diagnostics
48V bus · Isolated DC/DC · Protection · Monitoring
Feeds compute, sensors and actuator drivesMaps sensor timestamps into the real-time control domain without treating both networks as one clock.
Implemented in the RT host or a dedicated timing gatewayE-stop · Limits · Watchdogs · Fault handling
Project-defined controlled stop, brake / hold and drive inhibitTiming and sensing platforms
AI perception and planning.
Synchronized sensor acquisition.
Time-aware network connectivity.
EtherCAT inertial-state feedback.
Motion-control host; requires a real-time OS, control stack and cycle validation.
Clock-domain note: Native / PTP-timestamped perception sensors remain on the sensor-data plane. TALO-M1000 is shown only on EtherCAT DC and is not treated as a PTP-backbone sensor.
03
Head Intelligence Architecture
A synchronized head subsystem converts vision, audio and head state into local perception, natural interaction and coordinated robot responses.
Sensor data / system linkSynchronized streamsNeck actuationIndependent safetyPower distribution
01
Input
- Vision
- Front and peripheral cameras
- Audio
- Microphone array
- Head state
- IMU · Neck encoders
02
Synchronize
- Sensor capture
- CSI-2 · I²S / PDM · SPI
- Time alignment
- PTP / gPTP reference
- Event packaging
- Timestamp buffer · DMA
03
Understand
- Compute envelope
- Jetson Orin NX-class embedded module
- Scene understanding
- Multi-camera perception
- Speech & interaction
- Voice · intent · emotion
04
Respond
- Neck motion
- Yaw / pitch targets
- Interaction output
- Gaze · voice · expression state
- Robot coordination
- Body AI commands and telemetry
External interfaces
Independent infrastructure connections
48V input · Isolated DC/DC rails
Feeds · Entire head moduleLimits · Permission · Controlled stop request
Constrains · Neck motion10GbE · ROS 2 · Commands ↔ status
Coordinates · Response & System LinkHead sensing and robot-backbone integration
Synchronized sensor acquisition.
Time-aware network connectivity.
Placement note: TALO-B2000 is a PCIe GMSL2 capture card and belongs in a compatible host / bridge chassis, not at the raw sensor node or inside the head by default.
04
Dexterous Manipulation Architecture
A synchronized hand subsystem turns vision, touch and joint position into deterministic finger motion, while independent protection remains outside the sensor and cyclic-control device.
Sensor data · 10GbE + PTP-gPTP
Commands
Telemetry / status (return)
EtherCAT DC
Independent safety
ROS 2 · Policy inference · Kinematic targets
01
Perception
- Vision
- Palm camera
- Touch
- Fingertip pressure sensing
- Joint position
- Finger encoders
02
Deterministic control
Hand Control Core
CSI-2 · ADC / SPI
Encoder capture
Timestamp alignment
Deterministic events
Cyclic position / torque loops
Trajectory interpolation
Watchdogs · Current
Temperature · Fault status
03
Actuation
- Motor drive
- Multi-axis current control
- Finger joints
- MCP · PIP · DIP
- Wrist interface
- Power · brake · external I/O
Robot networks & independent safety
Deterministic wrist and joint execution
High-rate vision, telemetry and synchronization
Limits · Stop request · Brake / hold · Drive inhibit
Manipulation compute and connectivity
AI perception and planning.
Synchronized sensor acquisition.
Time-aware network connectivity.
Safety scope: Diagnostics inside the FPGA / control core do not constitute independent safety. A separate safety controller and safety I/O must enforce the project-defined safe state and be validated against the applicable machinery and drive-safety standards.
Need finger drives, actuators or a safety design? Discuss the mechanism, bus and certification requirements with our engineering team.Contact us →Build for the physical world