WIM Robot Controller HW+SW All-in-One Industrial AI Robot Controller

  • Compute, real-time control, and dev environment in one box
  • One controller drives up to 4 robots
  • 275 TOPS compute · 1ms (1kHz) real-time control
  • 24-axis simultaneous control · EtherCAT / EtherNet/IP
Request Quote
WIM Robot Controller AI Robot Controller
ONE JETSON · PHYSICAL AI

The board that runs your AI also moves the robot at 1 kHz

WIM Robot Controller runs the real-time control loop and the EtherCAT master inside the same NVIDIA Jetson that runs perception, reasoning and planning. Motion commands from the AI are taken on the same board and the built-in EtherCAT master drives the joints at 1 kHz, so there is no separate control PC, motion controller or EtherCAT master unit.

Task

Cameras · language

What it sees and hears becomes the task

WIM Robot Controller

One NVIDIA Jetson

GPU · AI

Perceive · decide · plan

VLA policy and vision inference, 10–30 Hz

Shared memory no copies

Isolated CPU cores · PLEM

Real-time control, 1 kHz

Kinematics, dynamics compensation, torque control, safe stop

EtherCAT master · PLEM

EtherCAT master built in

Drives CiA 402 drives directly, synced to the microsecond

While AI inference runs, the control loop still fires on time every 1 ms.

Robot

Drives · joints

EtherCAT servos, arms, humanoids, cells

From task to joint: AI policy → 1 kHz real-time control → EtherCAT master → drives. All of it, EtherCAT master included, runs inside one Jetson.

Today

An AI computer, a control PC, a motion controller and an EtherCAT master: 3 to 5 computers to buy and get talking to each other.

With WIM

One Jetson. AI inference, real-time control and the EtherCAT master run on the same board.

358 → 15 µs

Real-time latency, 20× steadier

Measured on Jetson AGX Orin

±4.1 µs

Cycle jitter, 12 axes at 1 kHz

Third-party test LTA LR500172603E

328 → 3.5 µs

EtherCAT sync deviation

Engineering blog 2026-10-10

6–12 mo → 3 days

Robot bring-up time

WIM 3-day bring-up

Milestones

  • NVIDIA Inception Grand Challenge 2026 Top 10

    NVIDIA Inception member, robot controller built on Jetson

  • Third-party test reports

    LTA LR500172603E (12-axis jitter at 1 kHz), KOTCA KST-25-236 (control-loop latency and periodicity)

  • Intel · Qualcomm · Kawasaki Robotics

    NDAs signed, physical AI discussions under way

  • 30+ sites

    Sites running controllers delivered by WIM

Key features

  • AI and control on one board

    AI inference, 1 kHz real-time control and the EtherCAT master run on the same Jetson. No control PC, no motion controller.

  • EtherCAT master built in

    The EtherCAT master lives inside the controller. It drives CiA 402 servo drives in position, velocity and torque modes, mixed on one bus.

  • Open down to torque control

    Gravity, friction and Coriolis compensation are built in, and your own torque controller plugs in through C++.

  • Four robots on one controller

    24 axes controlled at once, verified by driving four 6-axis robots.

  • Connects straight to ROS 2

    Supports ROS 2 Humble and Jazzy. Develop in C++, Python or over the REST API.

  • Open source, public docs

    The PLEM drive layer is on GitHub under Apache-2.0, and the documentation is open to everyone.

NVIDIA and Jetson are trademarks of NVIDIA Corporation. EtherCAT® is a registered trademark and patented technology, licensed by Beckhoff Automation GmbH, Germany. ROS 2 is a trademark of Open Source Robotics Foundation. Intel, Qualcomm and Kawasaki are trademarks of their respective owners.

Why are traditional robot controllers not enough?

01

Hardware Separation

AI compute (GPU) and real-time control are physically separated → communication latency, synchronization issues

02

Software Fragmentation

Different OS·drivers per robot type/chipset → increased development costs

03

Complex Systems

Separate wiring, external PC required → constraints on mobile/compact robots

04

Scattered Maintenance

Multiple devices and software vendors managed separately → slow root-cause diagnosis when failures occur

One Box Solves Everything

Traditionally, you had to buy hardware, install software, configure drivers, and set up communication separately. With WIM Robot Controller, everything comes in a single box.

HARDWARE

NVIDIA Jetson Orin AGX Controller

275 TOPS compute with 24-axis simultaneous control in a single module. Supports EtherCAT, EtherNet/IP, and Modbus TCP for instant integration.

BASED ON JETSON ORIN

275TOPS

Compute

24axes

Simultaneous

1ms

Control Cycle

<11kg

Weight

EtherCAT

Industrial RT

Ethernet

1Gbps

Wi-Fi · BT

Wireless

Digital I/O 16-port

In 8 / Out 8 · EtherCAT-expandable

APPLICATION
Pick & PlaceWeldingGrinding
AI · MOTION PLANNING
VLAVisionMotion Planning
Optimized and guaranteed by WIMFrom middleware down to the RT kernel, modified by us
MIDDLEWARE · WIM OPTIMIZED
ROS 2DDSEtherCAT Master
LOW-LEVEL CONTROL · WIM
Torque·PID·ADRCGravity·Coriolis·FrictionKinematics·Calibration
RT KERNEL · WIM TUNED
PREEMPT_RTCPU IsolationCUDA·TensorRT
HARDWARE
Orin AGXARM CPUGPU
SOFTWARE

PLEM — the ROS 2 robot SW platform, optimized by WIM

ROS 2, DDS, the EtherCAT master and the Preempt RT kernel are modified, optimized for Jetson and guaranteed by WIM. cuMotion motion planning comes pre-installed, so you can start developing in C++/Python right out of the box.

INTEGRATION

What Full-Stack Integration Really Means

System on Chip (SoC) eliminates CPU-GPU communication bottlenecks, while RT processes isolate CPU cores to guarantee 1ms (1kHz) periodicity without interference. Software that knows the hardware, hardware optimized for the software.

One Chip, Two Roles

AI and control run simultaneously

CPU
Unified Bus
GPU

Unified Memory

Zero-copy GPU ↔ CPU

CPU

RT processes isolate CPU cores for guaranteed 1ms periodicity

GPU

AI inference and vision processed in real-time on one GPU

PLEM

Robot Software Platform

PLEM

The middleware, low-level control and RT kernel that turn AI intent into smooth, precise motion are all modified and shipped by WIM. Hover a layer to explore.

Software Components

{}>

Robot Control Core Libraries

Core library (trajectory, position/torque control), Hardware communication library, Upper/lower integration interface, Sensor calibration, Customizable control algorithms

AI

AI / Physical AI Libraries

Calibration, CUDA/TensorRT optimization, Conveyor tracking

<<>>

Industrial Communication Drivers

EtherCAT Master, I/O Control, Multi-axis synchronization

~/>

Developer Tools & API

REST API, CLI tools, Logging, Monitoring dashboard

Built on ROS 2, PLEM adds the value

Go beyond the limits of ROS 2 — a production-ready development environment for industrial deployment.

Real-time Control

ROS 2 (Base)

Requires separate RTOS setup

+ PLEM

✓Native industrial real-time control

AI Integration

ROS 2 (Base)

Install external packages individually

+ PLEM

✓CUDA/TensorRT built-in

Deployment

ROS 2 (Base)

colcon build + dependency management

+ PLEM

✓One-click Docker container deployment

Hardware Abstraction

ROS 2 (Base)

Develop HAL from scratch

+ PLEM

✓Standardized abstraction API

Initial Setup

ROS 2 (Base)

Days for environment setup

+ PLEM

✓Start developing in 5 minutes

Industrial Communication

ROS 2 (Base)

Additional packages required

+ PLEM

✓EtherCAT Master built-in

Technical Specifications

HARDWARE

SoCNVIDIA Jetson Orin AGX
Compute275 TOPS (INT8)
Memory64GB LPDDR5 (Unified Memory)
Storage64GB eMMC + NVMe SSD
GPU ComputeCUDA 12.x
CommunicationEtherCAT, EtherNet/IP, Modbus TCP
Control Axes24-axis simultaneous
Control Cycle1ms (1kHz) / Jitter ±4.1μs
Emergency Stopmax 16.32ms (KIRIA certified)
I/O PortsDigital 8 in / 8 out (EtherCAT-expandable)
Network5x 1Gbps Ethernet, Wi-Fi, Bluetooth
Weight<11kg
Power90–264 VAC (universal)

SOFTWARE

SW PlatformPLEM (ROS 2, built-in)
OSPreempt RT Kernel
ContainerDocker Support
LanguagesC++ (primary), Python (extension)
Motion PlanningNVIDIA cuMotion
Control FunctionsMultiple control modes (position / torque / path interpolation)
AI LibrariesCalibration, Conveyor tracking, CUDA
UpdatesOTA remote update

Key Functions

Pick-and-place Trajectory

  • Section-based movement, speed control (acceleration)
  • Pick-and-place movement curve r-value change
  • Cubic spline trajectory
  • Conveyor tracking
  • Position control, Torque control
  • Place home designation (homing)
  • Coupling / Decoupling

Input/Output Control

  • Conveyor encoder control
  • Solenoid valve control

Calibration

  • Robot Calibration
  • Vision-robot calibration

Everything you need for AI robot control in a single box

Traditional (Separated)
WIM Robot Controller (Integrated)

Control Dev Time

EtherCAT setup, servo tuning, real-time guarantees yourself → 6 months~2 years

✓Built-in — works out of the box, dramatically shorter time to market

Control Stability

OS interference destabilizes control loop → precision work at risk

✓1kHz deterministic real-time control — applied in manufacturing & precision tasks, expanding into aerospace

System Complexity

Compute, control, and communication separated → many debug points, hard to isolate failures

✓Single box integration — fewer failure points, simpler maintenance

Dev Risk

Different env per manufacturer → relearn on every robot change

✓Standard environment — code reuse, portable across robots

Technical Assets

Black-box controller → customer know-how locked to vendor

✓Code & tuning 100% customer-owned — technology internalization

Find the Right WIM Robot Controller Configuration

Check what applies to you. Our engineering team will propose the optimal setup.

Real-World Applications

Waste Sorting

Waste Sorting

AI vision-based material and shape classification with automated sorting

Crop Harvesting

Crop Harvesting

AI ripeness detection for autonomous harvesting in unstructured environments

Parts Machining

Parts Machining

Precision position control with automated real-time quality inspection

Multi-Robot Simulation

Multi-Robot Simulation

Multi-robot cooperative control simulation and optimization

Frequently Asked Questions

Is your company a hardware or software company? What are your core technologies?+

We possess full-stack SW technology for AI robot development, including SoC-based RTOS development. (Full-Stack)

We have the technology to build a dedicated real-time operating system (RTOS) that ensures precise motion on the SoC — the 'brain' of the robot controller — along with all A-to-Z software technologies needed to make robots Move, See (Vision), and Think (AI).

We sell our technology in a form that customers can easily adopt.

We provide our full-stack technology so customers can easily adopt it. Customers only need to develop their core AI solutions on top of our platform, dramatically reducing development time and cost.

Different SoC manufacturers have pros and cons. Are you limited to a specific chipset?+

We currently use NVIDIA chips, chosen for GPU parallel processing capabilities.

We adopted the NVIDIA Jetson series as the proven hardware that can most quickly build the general-purpose GPU-based parallel processing environment needed for AI robot development.

We plan to expand to various chipset manufacturers including Samsung Exynos.

AI robots serve diverse purposes depending on their objectives and form factors, requiring the flexibility to choose appropriate chipsets. We are pursuing the business direction of developing our software to run on any manufacturer's chipset, including Samsung Exynos with its strengths in high-speed communication and low power consumption.

How does WIM Robot Controller differentiate from using NVIDIA Jetson standalone?+

Jetson is excellent 'compute hardware,' but it cannot become a 'real-time robot controller' on its own.

Jetson's default OS does not guarantee real-time performance, making precise motor control impossible. Most companies end up adding external controllers, which reintroduces data bottlenecks and cost increases.

We provide a full-stack solution that performs both AI computation and real-time robot control simultaneously on the high-performance Jetson SoC.

What advantages does WIM Robot Controller offer over traditional approaches?+

Technology internalization without depending on external solutions.

With internalized technology, you can reduce maintenance vendor costs or perform maintenance in-house.

Develop optimal solutions using your own data.

Nobody understands and manages your data better than you. Developing AI solutions in-house and applying them directly to your robots is the best way to achieve peak performance.

How is this different from conventional approaches?+

Previously, robots didn't have built-in GPUs

Traditional approaches require externally attaching GPUs to controllers, which demands extensive work on GPU drivers, AI SW libraries, frameworks, and robot communication. Our controller has GPU built into the hardware, and we provide the complete development environment so customers can focus solely on core logic development.

Previously, each robot brand could only be controlled by its own proprietary controller

Our controller works with any ROS2-supported robot or any motor with EtherCAT/CAN communication, offering true universality. As more robot companies adopt ROS2, a single WIM Robot Controller can control diverse robots from different manufacturers.

Most robot companies only allowed simple script modifications, not code-level access

We open everything down to the code level, providing developers with maximum freedom.

What does 1kHz real-time control actually mean on the shop floor?+

Deterministic 1 ms control cycle (±4.1 μs jitter)

Motor commands are issued every 1 ms with ±4.1 μs jitter (third-party tested), preserving path accuracy in high-speed machining, precision assembly, and vision-tracking tasks.

Stability vs. general-purpose OS

Generic Linux/Windows-based controllers suffer scheduler interference that destabilizes the control loop and puts precision work at risk; W-RC guarantees cycle time via a Preempt RT kernel architecture.

What does 275 TOPS actually mean for on-device AI inference?+

Run vision and control networks concurrently

With 275 TOPS (INT8) of compute, multiple networks — object detection, pose estimation, trajectory generation — run concurrently on-device without an external server.

Edge inference removes network latency

Inference and motor control share one box, eliminating round-trip latency to external GPU servers and enabling real-time closed-loop AI control.

Start with Physical AI

Request QuoteStart developing on PLEM Docs