AN-001 · Embedded & Robotics · In progress
Started Jul 2026

A $3,000 haptic controller became a $300 board inside the device

1.0Abstract

At CMU’s Microdynamic Systems Laboratory I replaced the desktop-sized controller of a 6-DOF magnetic levitation haptic device (a legacy QNX machine, PCI DAQ cards and linear amplifiers in a full ATX case) with an STM32H7 board that mounts inside the device. That meant porting the firmware off QNX, driving AD7606 and LTC2688 converters over SPI with DMA within a 15 µs budget, and redesigning the amplifier as a modular unit. I also wrote a driver so the device’s force feedback works inside the SOFA simulator. Driving the lab’s ballbot with it is written up separately under Shmoobot.

At a glance

Parameter	Value
Status	In progress. Prototype, not production
When and where	Jul 2026 – present, Microdynamic Systems Laboratory, Carnegie Mellon
My part	Undergraduate researcher: the controller, its firmware, the amplifier redesign, and the SOFA integration
Built with	STM32H7; AD7606 ADC and LTC2688 DAC over SPI with DMA; a DRV8262 driver board in KiCad; SOFA
Before	A legacy QNX real-time PC with PCI DAQ cards and linear amplifiers, in a full ATX case: about $3,000
After	A ~$300 STM32H7 board that mounts inside the device, with ~2× the compute and a 15 µs sensor-to-actuator budget
Amplifier	Redesigned as a modular unit inside the device: a 7-channel DRV8262 board with INA240 current sensing, designed in KiCad and not yet fabricated; a switching PWM stage in progress
On show	Robotics & AI Discovery Day, Sep 16 2026; IROS 2026, Sep 27 – Oct 1, at the Butterfly Haptics table

A lab bench: an STM32 Nucleo-H753ZI board wired with jumper leads to a breadboard, two DB37 breakout boards and an Analog Devices evaluation board, with a Rigol bench power supply behind them — The controller as a bench prototype, August 2026: an STM32H7 Nucleo board, an Analog Devices converter evaluation board and DB37 breakouts, on a bench supply.

The setup

The lab’s maglev haptic device floats a handle on magnetic fields and pushes back on your hand with whatever force the software asks for, in six degrees of freedom. The controller that closed that loop lived in a full ATX PC case: a legacy QNX real-time machine, PCI DAQ hardware, and linear amplifiers. One note on the photographs: the controller design belongs to the lab, so they show the bench, the device and the demos rather than the board in detail.

A lab room with two desks: on the left a monitor above an orange-and-black maglev haptic device and a keyboard, PC towers on the floor between the desks, and posters about maglev haptic interfaces on the wall — The bench at MSL, August 2026. The orange dome on the left desk is a maglev haptic device.

The controller moved inside the device

The main result is that the whole control stack, a $3,000 desktop-sized controller, became a $300 board that mounts inside the haptic device, as a modular design that can be swapped and extended. The size collapse is real but unmeasured, so this page describes the change rather than claiming a multiplier.

   position sensors (6-DOF maglev device)
              │
              ▼                       ┐
   AD7606 ADC ─ SPI with DMA          │
              │                       │
              ▼                       │  15 µs
   STM32H7 (Cortex-M) ─ servo loop,   │  sensor-to-
   firmware ported off QNX            │  actuator
              │                       │  budget
              ▼                       │
   LTC2688 DAC ─ SPI with DMA         │
              │                       │
              ▼                       │
   amplifier ─ redesigned, modular,   │
   inside the device                  ┘
              │
              ▼
   flotor coils → handle → sensors

Why the budget is fifteen microseconds

A magnetically levitated flotor is open-loop unstable: it falls unless the loop catches it. That is what the sensor-to-actuator budget buys, and it is why the number is a hard constraint rather than a target. Drag the latency below and watch the loop lose it. The model loses the flotor at roughly 110 µs; the real device has six coupled axes, sensor noise and amplifier bandwidth that the model leaves out, which is why the budget sits well inside that edge.

Disturbance response, 0.2 mm initial displacement

Loop latency 15 µs

Verdict

—

Settling (±10 µm)

—

Peak overshoot

—

Margin to instability

—

Closed-loop disturbance response against control latency, 0.2 mm initial displacement.

This is an idealized single-axis model with delayed state feedback. It shows why latency matters; it is not a measurement of the device, and no loop figure on this page is measured yet.

Firmware

I ported the firmware off the legacy QNX machine onto the STM32H7, replacing the PCI DAQ path with an AD7606 ADC and an LTC2688 DAC driven over SPI with DMA, and budgeted a 15 µs sensor-to-actuator path against the legacy loop. The STM32 gives roughly 2× the compute throughput of the QNX software.

The amplifier

The linear amplifiers went too. I redesigned the power amplifier as a modular unit that lives inside the device, and designed a 7-channel DRV8262 motor-driver PCB in KiCad with INA240 current sensing and 2512 shunt resistors as the replacement. It is designed but not yet fabricated. A switching PWM-mode amplifier for better efficiency and a smaller footprint is in progress. An earlier interim step used a DB37 breakout board to drive the original amplifier while the rest of the stack was being moved.

Haptics in the simulator

The device is programmed against the MLHI API. I wrote a driver that exposes MLHI to SOFA, a separate simulation framework, so haptic force feedback runs inside the simulation loop (you can feel the simulated scene), updated 8 demos, and debugged the SOFA haptic drivers for instrument velocity in collision response and rotational force output.

A monitor showing a SOFA simulation scene, a shaded box with a grey instrument inside it, beside a maglev haptic device with a hand resting on its handle — Feeling a simulated scene: SOFA on the monitor, the device under the hand. A frame from the demo video.

The ballbot

The same device later drove the lab’s ballbot, the Shmoobot, with force feedback, so the operator’s hand gets information back from the robot instead of only sending commands to it. That work is written up separately under Shmoobot.

What’s next

The switching PWM amplifier, and fabricating the DRV8262 board. In September I’m demoing and presenting the haptic-device work at CMU’s Robotics & AI Discovery Day (Sep 16), and demoing the SOFA integration and demos at the Butterfly Haptics table at IROS 2026 (Sep 27 – Oct 1).

← All projects