Compute & control

Runs perception, planning and control workloads. General computing hardware and real-time motor control serve different roles in the stack.

Inside the compute & control assembly

Processing, joint control, power distribution and cooling share the robot’s electronics bay. Open the assembly to inspect each role.

Drag to rotate · scroll to zoom

Loading model…

Inside the compute & control assembly

Select a part to see its role.

Source & model scope

Illustrative compute and control architecture. Board layout, connectors and mounting are schematic, not a specific robot’s hardware. Original Korthos geometry.

NVIDIA · module and thermal design documentation · Model scope & reuse

Representative products

Component directory
Architecture, materials & specificationsArchitecture options, material requirements and integration constraints

Architecture options

Integration format and control architecture are separate choices. A robot can combine a compute module with dedicated joint controllers.

ApproachMechanismAdvantagesTrade-offs
Compute moduleProcessor and memory connect to a carrier board designed for the robot.Lets connectors, dimensions and cooling fit the body.Requires carrier-board design, power integration and thermal validation.
Developer kitA ready-made board packages the module with evaluation hardware.Speeds software development and sensor testing.Size, connectors and cooling may be unsuitable for a finished robot.
Separate real-time controllersDedicated controllers handle time-critical motor commands.Keeps joint control cycles separate from variable AI workloads.Adds communication, timing and fault-handling work between controllers.

Specifications & operating limits

TOPS measures arithmetic throughput under specified conditions. Memory capacity limits what can fit, while bandwidth affects data movement. Workload and power setting determine the resulting latency.

Materials & constraints

Semiconductor packages
Provide processing, memory and power management.
Copper PCB traces
Connect the components and carry electrical signals.
Thermal interfaces & heatsinks
Transfer heat from the processor into the cooling system.

Dependencies & engineering limits

Memory supply & module lifetime
A compute module depends on memory, storage and power-management parts with their own lifecycles. Replacing a discontinued part can require board changes, software updates and system testing.
AI compute & motor control
AI processors handle perception and planning; real-time controllers regulate motion using current and position feedback. Controller count depends on the architecture: TI’s TIDA-010992 reference design controls six hand motors with one C2000 MCU. This is a reference-board configuration.
Integrated actuators
Power & cooling & sustained speed
When temperature or power limits are reached, a processor may reduce its clock speed. Available cooling sets the sustainable workload inside the robot.
Batteries & power
Board integration & testing
A module still needs suitable power rails, data interfaces and cooling. Concurrent sensors and workloads share bandwidth, power and cooling, so their combined behaviour can differ from isolated benchmarks.
Model size & memory demand
Memory holds both the model and its working data. Data transfer can limit throughput before the processor reaches its arithmetic capacity.
Communication & control timing
EtherCAT, CAN-FD and Ethernet interfaces serve different control architectures. Protocol support, bus load and command/feedback timing determine the delay between the controller and joints.
Integrated actuators

Manufacturing & assembly

Guide

About this route

The route shows one process or architecture. Select a step to see its function and connections; related parts can be parallel inputs or alternatives.

Function

Inputs & dependencies
Outputs & connections
Related processes & parts
Sources

Integration / calibration · Representative route. Connect Tech’s Forge manual provides the integration example.

Related parts & processes 6

Compute module

A module packages core compute functions for integration into a larger system.

Module interfaces and supported software constrain interchangeability.

Sources & technical scope

Illustrative images of generic materials and components.

Related references
Link to this detail

Production equipment

Market context

Jetson production-module pricing

US$999Orin NX 16GB · module · 1,000+ units
US$2,999AGX Orin 64GB · module · 1,000+ units
US$4,999Thor T5000 · module · 1,000+ units

A dated benchmark for onboard compute procurement. Module prices exclude the rest of a robot computer, including its carrier, cooling and enclosure.

ANYbotics

Zollner manufactures ANYmal’s PCBs, pre-assembles modules and completes robot assembly in Hombrechtikon, Switzerland. It also handles purchasing for more than 400 components, with specifications updated between production batches.

Jetson

NVIDIA lists Orin NX and commercial AGX Orin modules as available through January 2032, Thor T5000 through August 2035 and Thor T4000 through January 2036.

Unitree

Go2 EDU adds a depth camera and supports a separate high-compute module, including Orin options. The standard onboard eight-core CPU remains a distinct computing layer.

Explore documented supply relationships →