ADI Robotics: Tactile Sensing and Precision Control

A robot is a closed-loop system, not a collection of AI chips #

Humanoid robots and dexterous hands are often discussed as an artificial-intelligence problem. In production hardware, however, the difficult work happens between an inference result and a safe, repeatable physical action. The machine must detect contact, estimate pose, move a joint, measure the result and correct the next command within a tightly controlled time budget.

That makes Analog Devices (ADI) relevant beyond its better-known converter and amplifier franchises. ADI positions its robotics portfolio around sensing, real-time signal processing, deterministic connectivity, precision motion and functionally safe power control. The useful procurement question is therefore not “Which ADI chip runs the robot?” but “Which signal-chain function needs to be solved, and what evidence supports each component choice?”

This guide maps concrete product families and orderable models into that loop. It also preserves an important boundary: ADI has publicly demonstrated multimodal tactile technology under development, but that does not mean every tactile function shown in a concept diagram already has a publicly orderable sensor SKU.

Where ADI fits in the robot stack #

ADI’s public robotics material describes intelligence distributed from a central planner down to joints, limbs and fingertips. That architecture is practical because raw sensor data is expensive to move, while motor and safety loops cannot always wait for a central processor.

The table is deliberately not presented as a validated reference-design BOM. Electrical compatibility, application positioning and a vendor-verified complete robot design are three different evidence levels.

Perception: depth, inertial state and joint angle #

ADTF3175 for depth rather than conventional imaging #

The ADTF3175 is a 1-megapixel indirect time-of-flight module. ADI integrates the image sensor, optics, optical band-pass filter, infrared illumination and laser driver in the module. The official product page specifies a 1024 × 1024 depth image, a 75° × 75° field of view and an operating range from 0.4 m to 4 m at 15% reflectivity. Its quoted ±3 mm depth error belongs to the vendor’s stated test conditions; it should not be generalized to every robot enclosure, target surface or ambient-light environment.

For robotics, the module can support near-field ranging, obstacle perception or coarse object geometry. It does not replace RGB vision, tactile sensing or safety-rated detection. Optical window material, eye-safety integration, multipath interference and calibration after mechanical assembly remain system responsibilities.

ADXL355 for low-drift motion measurements #

The ADXL355 is a low-noise, low-drift, three-axis digital accelerometer with selectable ±2 g, ±4 g and ±8 g ranges. ADI lists robotics, platform stabilization and tilt sensing among its applications. Its 20-bit ADC, SPI/I²C interfaces, integrated temperature sensor and specified −40°C to +125°C operating range make it useful where repeatable low-frequency motion or structural movement matters more than consumer-grade feature density.

It is not a complete six-axis IMU. A robot that needs angular rate, synchronized multi-axis timing or calibrated orientation output should evaluate an ADIS inertial measurement unit or another integrated IMU rather than treating an accelerometer as a drop-in substitute.

ADA4571 and ADMT4000 for different position problems #

The ADA4571 provides sine and cosine outputs from an anisotropic magnetoresistive angle sensor. ADI specifies 180° sensing, a maximum angular error of 0.5° and operation from −40°C to +150°C. It is a plausible choice for compact motor or actuator commutation when the magnet geometry and calibration budget fit.

The ADMT4000 addresses a different problem: retaining absolute multiturn position through power loss without a backup supply. That can matter in joints, grippers or linear mechanisms where the controller must know mechanical state immediately after startup. Procurement teams should not collapse “angle sensor” and “multiturn absolute position” into one interchangeable category; the magnet, shaft arrangement, required turns, interface and safety concept determine the correct family.

Tactile sensing: the most important evidence boundary #

ADI’s recent robotics material describes a multimodal tactile sensor prototype intended to capture cues such as force, torque, slip or material-related information. The engineering direction is meaningful: vision can lose contact information when an object is occluded by a hand, while compliant or transparent objects may be difficult to classify from images alone.

But the procurement interpretation must stay conservative. Public material describes the tactile technology as under development. It should not be converted into an invented orderable part number, guaranteed production date or claimed qualification level.

The ADPD7000 is an orderable multimodal sensor front end, and it can synchronously acquire several sensor modalities. That makes it relevant when evaluating a custom optical, electrical or electrochemical sensor stack. However, its public product page does not establish it as the commercial version of ADI’s robotics tactile prototype. In this article it is therefore a signal-chain candidate, not an official tactile-hand reference design.

For a real fingertip or skin PCB, the RFQ and engineering package should define:

• transducer type and full-scale signal range;
• required simultaneous channels and sampling rate;
• allowable end-to-end contact latency;
• noise floor after flex cable, connector and mechanical integration;
• local preprocessing or compression requirements;
• calibration method across temperature and elastomer ageing;
• expected flex cycles, ingress protection and replaceability.

The sensor element, analog front end, mechanics and inference model must be validated as one assembly. Choosing a high-resolution ADC alone does not create useful touch.

Precision motion: from torque command to measured joint position #

TMC4671 and TMC6200 for compact servo axes #

The TMC4671 implements field-oriented control in hardware for brushless DC, permanent-magnet synchronous, two-phase stepper, DC and voice-coil motors. It supports torque, velocity and position control and integrates ADC and position-interface functions. Moving the fast control loop into dedicated hardware can reduce software timing variability in multi-axis machines.

The TMC6200 is the companion high-power gate driver for external MOSFETs. Together, the pair can form the control and drive core of a compact servo axis, but they do not eliminate power-stage engineering. MOSFET selection, shunt layout, dead time, gate resistance, current range, bus transients, thermal path and electromagnetic compatibility still determine whether the actuator survives its load profile.

The correct sourcing unit is the full orderable code—such as TMC4671-LA—not only the family name. Confirm package, reel or tray format, temperature grade and revision before releasing an approved BOM.

AD2S1210 where a resolver earns its complexity #

For harsh-duty joints or large servo axes, a resolver may offer mechanical and environmental robustness that a small magnetic encoder cannot. The AD2S1210 combines resolver excitation with 10- to 16-bit position conversion, velocity output and programmable fault detection. ADI lists DC and AC servo motor control among its applications.

This is not the default answer for every humanoid joint. Resolver size, excitation circuitry and cost can be excessive for a compact finger actuator. It belongs in the option set when temperature, contamination, vibration or functional-safety requirements justify it.

Current sensing closes the torque loop #

A servo’s torque estimate commonly depends on phase-current measurement. ADI’s AD8418/AD8418A zero-drift, bidirectional current-sense family is one possible signal-chain element, while isolated modulators or simultaneous-sampling ADCs may fit higher-voltage architectures. The choice depends on common-mode voltage, PWM edge rejection, shunt dissipation, bandwidth, isolation boundary and the number of axes sampled together.

Treat current-sense gain and shunt value as system parameters. A substitute with the same package but different input filtering, offset behavior or PWM rejection can destabilize calibration even when the board powers up.

Connectivity and edge compute #

Robots need different networks at different layers. Camera streams require high bandwidth; joint telemetry may require modest bandwidth with predictable latency; safety traffic may need an independently justified architecture.

The ADIN1110 integrates a 10BASE-T1L Ethernet MAC-PHY and supports 10 Mbit/s communication over a single twisted pair, with diagnostics and IEEE 1588 timestamping features. It can reduce host interface burden in distributed industrial sensor nodes. Whether it belongs inside a particular robot depends on cable length, connector mass, topology, cycle time and protocol overhead. It should not be presented as the universal replacement for EtherCAT, CAN FD, 10BASE-T1S or proprietary joint links.

For very local inference, the MAX78000 combines a microcontroller with a low-power convolutional-neural-network accelerator. It is suited to bounded tasks such as keyword, anomaly or compact visual classification when the model fits its supported memory and operators. It is not a substitute for a central NVIDIA Jetson Thor-class planner or a large vision-language-action model.

That separation—small deterministic local loops plus powerful central planning—is consistent with ADI’s published “physical intelligence at the edge” direction.

Safety and energy cannot be bolted on later #

Robot safety is a system claim, not a property inherited from one component. Fault detection in a resolver converter, watchdogs in a controller or diagnostics in an Ethernet PHY can contribute evidence, but they do not make the complete robot functionally safe.

Likewise, battery monitoring devices such as the LTC6813-1 can monitor multi-cell stacks and communicate through isoSPI, but cell count alone does not qualify a monitor for a mobile robot. Designers must define pack voltage, cell chemistry, measurement error, isolation, balancing strategy, fault containment, contactors, precharge and regulatory requirements.

The practical architecture should separate at least three timing domains:

1. Fast control: current and torque loops close locally. 2. Coordinated motion: joint position and trajectory commands move deterministically across the machine. 3. Planning and perception: central AI decides goals without becoming the only protection against an unsafe actuator command.

Procurement map: buy by function and evidence #

Before an RFQ, include the robot subsystem, exact manufacturer part number, target quantity, prototype versus mass-production phase, required date code, package format, operating temperature and any approved-source restrictions. For motor-control parts, also state bus voltage and motor type; for sensors, state range, bandwidth and interface.

A practical design and sourcing checklist #

• Confirm whether a product is explicitly positioned for robotics, merely electrically compatible, or part of a technology demonstration.
• Freeze complete suffixes and package codes; family names are insufficient for purchasing.
• Validate sensor and actuator latency as an end-to-end loop, not as isolated data-sheet numbers.
• Treat calibration data, firmware and model conversion tools as controlled BOM dependencies.
• Review lifecycle status and qualification evidence directly on the manufacturer’s current product page.
• Require lot traceability, packaging condition and moisture-sensitivity controls for production orders.
• Do not claim functional safety, tactile performance or robot-level reliability from a single component certificate.

Conclusion #

ADI’s robotics value is strongest when its products are treated as a connected signal chain: ADTF3175 and ADXL355-class devices observe the machine and environment; ADA4571 or ADMT4000-class sensors measure position; TMC4671 and TMC6200 drive compact servo axes; AD2S1210 supports resolver-based feedback; ADIN1110 and MAX78000 can distribute selected communication and inference functions.

The tactile story is strategically important, but procurement language must remain exact: ADI has described multimodal tactile technology in development, while the commercial component selection still depends on the final sensor architecture. A credible robot BOM therefore maps each function to an orderable code, labels its evidence level and verifies the complete loop before volume purchase.

Official references #

• ADI: Architecting Physically Intelligent Robotics at the Edge
• ADI: Industrial robotics solutions
• ADI adopts NVIDIA Jetson Thor for humanoid robotics
• ADI: The touch that will transform robotics
• ADTF3175 product page
• ADXL355 product page
• ADA4571 product page
• ADI magnetic position-sensor portfolio, including ADMT4000
• TMC4671 product page
• TMC6200 data sheet
• AD2S1210 product page
• ADIN1110 product page
• MAX78000 product page
• ADPD7000 product page
• LTC6813-1 product page