Versatility
Integrate different analog-output sensors through one centralized ADC.
Implemented in the hand20 channels16 joint + 4 tactile sensing channels
IEEE ROBOTICS AND AUTOMATION LETTERS · EARLY ACCESS 2026
High-Speed, Scalable Sensor Readout for Dexterous Robotic Hands via Shift-Register Multiplexing
1 Soft Robotics Lab, IRIS, D-MAVT, ETH Zurich 2 Integrated Systems Laboratory, ETH Zurich
Joint and tactile feedback, connected through one compact readout architecture.
Scalability
Extend sensor chains and select finger-level branches with dedicated input pulses.
Between sensor modules3 signal linesCLK, PULSE and Vout · excluding power and ground
Fast and configurable sampling
Set the full-scan rate through the clock. ADC conversion time chiefly limits the maximum rate.
Across all 20 channels1 kHzFull-scan rate · stable operation up to 1.5 kHz
Active motionWater-bottle interaction with live sensor readout.
Externally applied motionFollow the hand as its joints are moved by contact.
REPEATED INTERACTION
Watch what the
joint sensors see.
Repeated contact reveals the difference between directly measured joint angles and motor-based estimates.
Explore the experiments ↓THE IDEA
More sensing.
Less wiring complexity.
Dexterous hands need to know both how they move and what they touch.
Adding sensors inside a tendon-driven hand competes with tendons and joints for space. This work centralizes analog-to-digital conversion and uses a shift register to select sensor outputs in sequence over a shared analog line.
Each sensing unit adds only a D flip-flop and an analog switch to its base sensor. The same readout supports different analog-output sensors, serial expansion, and selective access to finger-level branches.
Read the paper abstract
Dexterous robotic hands require high-speed multimodal sensing across many degrees of freedom, yet existing readout architectures often impose trade-offs between sensor count, wiring complexity, and sampling bandwidth. This paper presents a scalable analog sensor readout architecture based on a serial-in parallel-out (SIPO) shift-register principle. The proposed architecture supports versatile integration of heterogeneous analog-output sensors, scalable expansion using only three signal lines between sensor modules, and fast, configurable sampling. We validate the approach on a tendon-driven robotic hand integrating 16 joint sensor modules and one four-channel tactile sensor module, enabling acquisition of 20 sensor channels at a full-scan rate of 1 kHz, with stable operation up to 1.5 kHz. Joint sensor characterization showed a maximum slope absolute percentage error (APE) of 0.446% and sub-degree estimation error, indicating that the proposed readout system preserves sensing performance. For tactile sensing, LSTM-based models enabled real-time force estimation and contact-location classification at 1 kHz. System-level experiments showed that the joint sensors provided more accurate feedback than motor-based estimation during interaction. Closed-loop experiments further demonstrated improved joint trajectory tracking using joint-sensor feedback and force-controlled object pinching using tactile feedback. The proposed architecture offers a practical path toward sensorized robotic hands for dexterous manipulation.
THE ARCHITECTURE
One pulse.
One sensor at a time.
A serial-in parallel-out shift register moves the selection pulse along the chain. Each selected switch connects its sensor to the shared ADC.
LIVE SIGNAL WALKTHROUGH
When PULSE OUT is high, the switch closes.
CIRCUIT
Switch openSensor isolated from Vout
SYNCHRONIZED TIMING0.00 T
ClockPulseAnalog signalADC samplingSensor disconnected
Follow the sequence, at your own pace.Select a step to pause at that moment.
Slowed, idealized illustration of the readout principle. T is one clock period. Analog voltage is illustrative, not experimental data. Vout is drawn only while this sensor is connected. Blank intervals indicate disconnection, not zero volts. The dashed line beside the switch indicates its control linkage. ADC reading is illustrated after the falling clock edge. Power and ground remain connected.
FROM ONE UNIT TO A SENSOR CHAIN
One branch. Many sensors. One output.
Follow one selection pulse along the chain. Each unit connects its sensor to the shared analog line for one clock period.
3 T per full scanFixed clock period T · one branch · centralized ADC
Waiting for the first pulseVout: disconnected0.00 T
ADC samplingVout recorded at the marked instant.
Try different sensor voltages.Live inputs · recorded history stays unchanged
A rising edge shifts the pulse to the next flip-flop. Only the selected unit closes its switch.
Vout follows that unit’s sensor voltage. Dots mark ADC reads after each falling clock edge.
A new input pulse starts the next scan. At a fixed clock rate, N units need N clock periods per full scan.
Idealized timing and example voltages, not oscilloscope measurements. Switching transients, ADC conversion time and electrical settling are omitted. Power and ground connections are omitted from the drawing. Blank Vout before the first selection means disconnected, not zero volts. Voltage changes apply only to newly drawn time; rewinding preserves recorded values. The PC receives one value per ADC sample, in unit order; -- means not yet received. It keeps the latest two scans across plot loops. Reset or changing the unit count clears the received data. Each plot loop starts a fresh waveform recording. With one unit, Q1 stays high between repeated scans because each rising edge captures a new input pulse.
HARDWARE & INTEGRATION
From sensor modules
to a sensorized hand.
Sixteen joint sensor modules and one four-channel tactile module connect through flexible interconnects to a shared readout.
One network, built into the hand.
Flexible interconnects route the joint and tactile sensor signals to the wrist-mounted interface board.
Joint position
An AS5600 magnetic rotary sensor measures the joint angle locally.
Fingertip touch
Four Hall-effect sensors feed separate LSTM models for force estimation and contact-location classification.
Shared conversion
A Teensy 4.1 selects sensor outputs, samples their voltages and sends the readings to a PC.
How the branches connect
The MCU supplies a shared clock and a separate selection pulse for each finger branch. Within a branch, the pulse advances one sensing unit per clock cycle, connecting its output to the shared ADC. The MCU sends the readings to the PC over USB serial for processing in ROS 2.
EXPERIMENTAL RESULTS
Measure the hand.
Close the loop.
Validation spans sensor characterization, physical interaction, joint trajectory tracking, and tactile-based pinching.
0.233–0.368°
Joint-angle estimation RMSE
Range of mean RMSE across seven conditions, with five trials per condition, using the reference sensor slope (Table I).
0.125 N
Tactile force estimation RMSE
Entire test set (Section III-B). For force above 0.05 N, RMSE is 0.148 N.
93.4%
Contact-location accuracy
Test-set accuracy over five classes: no contact and four predefined contact locations (Section III-B).
Characterize each sensing modality.
Joint sensing and fingertip force estimation, in motion.
Explore sensor characterization figures
Direct sensing during contact.
Tendon compliance and transmission uncertainty can make motor-based joint estimation less reliable during physical interaction. Dedicated joint sensors tracked vision-based ground truth more closely, even after additional calibration of the motor-based estimates.
| Joint | Joint sensor (ours) | Motor-based |
|---|---|---|
| MCP | 0.387° | 2.343° |
| PIP | 0.598° | 2.599° |
Better joint trajectory tracking.
Sensor feedback reduced mean tracking RMSE across 16 joints from 3.13° to 0.92° in 1 Hz sinusoidal trajectory experiments. Both controllers ran at 200 Hz.
Mean (standard deviation): open loop 3.13° (0.39°); closed loop 0.92° (0.24°). Tracking error was measured against joint-sensor readings.
A grasp guided by touch.
A rigid cube, a soft ball, and a paper cube: tactile feedback regulates pinching force across objects with different compliance.
Steady-state force tracking RMSE: 0.038 N, 0.032 N, and 0.055 N, respectively, measured between the unfiltered force estimates and target forces from 3.0 to 4.8 s.
View the closed-loop control results
FULL RESEARCH FILM
From readout to real interaction.
Research film · 3 min 21 s
Hardware integration, multiplexed sensing, experimental validation, and feedback control.
PAPER & RESOURCES
Explore the research.
Published October 8, 2026 · IEEE Robotics and Automation Letters · Early Access
DOI: 10.1109/LRA.2026.3741779 ↗
Citation
@article{kim2026sensorreadout,
title = {High-Speed, Scalable Sensor Readout for Dexterous
Robotic Hands via Shift-Register Multiplexing},
author = {Kim, Jaehoon and An, Boshi and Christoforidis, Lazaros
and Papadakis, Michail and Kartsch, Victor
and Katzschmann, Robert K.},
journal = {IEEE Robotics and Automation Letters},
year = {2026},
pages = {1--8},
doi = {10.1109/LRA.2026.3741779},
url = {https://ieeexplore.ieee.org/document/11726871},
note = {Early Access}
}IEEE Early Access citation · published October 8, 2026. Volume and issue are not yet assigned.
Acknowledgements
This work was supported by SNSF Project Grant #200021_215489. Michail Papadakis received scholarship support from the Eugenides Foundation.