Latest / Embodied AI 101 / Feeling Around Corners: STRP Compresses Fingertip Force into Directional Touch

Transcript

Feeling Around Corners: STRP Compresses Fingertip Force into Directional Touch · Embodied AI 101

• Feeling Around Corners: STRP Compresses Fingertip Force into Directional Touch. STRP: A Low-Cost Teleoperation Platform with Spatial Force Feedback for Visually Occluded Dexterous Manipulation in Hazardous Environments • Jingyuan Luo, Siquan Wu, Li Shen, Chuliang Chi, Hao Wu, and Yongquan Chen • Shenzhen Institute of Artificial Intelligence and Robotics for Society; The Chinese University of Hong Kong, Shenzhen; Shenzhen XJCSENSOR Technology Co., Ltd. • Biomimetics 11(9):609 • 2026. Imagine guiding a dexterous robotic hand through an opaque electrical cabinet. The camera can show you where the wrist entered, but not whether the little finger is brushing the right wall, pressing into the back of the enclosure, or approaching a fragile component from below. A conventional vibration motor might tell you that contact occurred. It does not tell you where around the fingertip that contact occurred. That is the narrow but important problem addressed by STRP, the Spatial Teleoperation Research Platform. The system measures forces at four robotic fingertips and converts the lateral components into vibrations at matching locations on the operator’s fingers: left, right, dorsal, and volar. Instead of attempting to recreate remote mechanics through a heavy force-reflecting exoskeleton, STRP turns contact direction into a compact cutaneous code. The paper, published on August 28, 2026, presents a complete prototype built around a Hall-sensor exoskeleton glove, a ZWHAND dexterous hand, miniature six-axis force–torque sensors, and a UR5 arm. Its most striking results include 97.5% identification accuracy for four cardinal tactile cues, an approximately 28-millisecond electronics and communication latency, a 41% reduction in beaker-placement time, and complete wall localization inside an opaque enclosure under directional feedback. But the most useful way to read this work is not as a finished hazardous-environment robot. It is better understood as a hardware hypothesis: perhaps high-value haptic teleoperation does not require physically reproducing every remote force. Perhaps it is enough to preserve the task-relevant structure of contact—especially which side of the fingertip is touching something. The Missing Variable Is Contact Direction. Most low-cost dexterous teleoperation systems are built around an essentially one-way information flow. The operator supplies hand pose, a retargeting system generates robot commands, and cameras return visual observations. If tactile feedback exists at all, it is often one vibration channel per finger. That channel answers a binary or scalar question: is the finger touching something, and perhaps how strongly?. Under good visibility, the operator can combine that signal with the image and infer the contact geometry. Under occlusion, those variables collapse together. A strong vibration could mean that the fingertip pad has established a stable grasp, that the side of the finger is scraping a wall, or that the back of the finger has become trapped beneath an object. STRP’s argument is that the sign and orientation of the contact force are useful information in their own right. A right-side contact should stimulate the operator’s right fingertip skin; a dorsal contact should stimulate the dorsal skin. The operator does not receive a numerically accurate reconstruction of the force vector. Instead, the remote vector is translated into a body-referenced tactile pattern. This distinction places STRP between simple contact alarms and true kinesthetic force reflection. Kinesthetic interfaces resist the operator’s motion and can approximate object impedance, weight, or joint torque. STRP cannot physically stop the operator from closing a finger. Its vibrations are sensory substitution: the operator must perceive the cue, interpret it, and voluntarily adjust the command. That trade is central to the low-cost design. Cutaneous actuators are small, light, and relatively easy to distribute around a finger. They also avoid many of the stability, backdrivability, and mechanical-transmission problems associated with bilateral force-reflecting exoskeletons. What is lost in physical fidelity may be recovered through a deliberately designed information code. The System from Human Joint to Robot Fingertip and Back. STRP contains two parallel loops. The forward loop captures the operator’s hand configuration and drives the robot. The return loop reads fingertip forces and drives the glove’s tactile actuators. The host computer coordinates the arm and hand through ROS 2, while the tactile path runs as a dedicated polling process. A deliberately simple master glove. The master interface is a linkage-based exoskeleton fabricated largely from 3D-printed PLA. It carries 15 Hall-sensor measurement channels, while five additional joint variables are computed through assumed interphalangeal coupling. Small radially magnetized neodymium magnets rotate with the joints, and side-mounted linear Hall sensors observe the changing magnetic field. The mechanical alignment keeps each joint’s operating range within a monotonic part of the sensor’s approximately sinusoidal response. That permits an intentionally simple calibration model rather than requiring a learned nonlinear estimator. Before operation, the wearer spends ten seconds repeatedly opening the hand and making a fist. The observed minimum and maximum voltages are then mapped to the slave hand’s command range. One interphalangeal angle can be inferred from another using an empirical coupling relationship. This design is inexpensive and easy to recalibrate after fit changes or modest mechanical drift. It is also closely tied to anthropomorphic joint correspondence. There is no sophisticated task-space retargeting layer described here—no optimization over fingertip poses, no contact-aware objective, and no explicit compensation for large embodiment differences. The approach is most natural when the human and robot hands have roughly corresponding digits and flexion patterns. The paper’s kinematic validation reflects that emphasis. The authors demonstrate four static configurations: full extension, isolated index extension, simultaneous index-and-little-finger extension, and a lateral pinch. The robot qualitatively reproduces all four without obvious inversion or coupling errors. There is no reported joint-angle RMSE, fingertip-position error, dynamic trajectory error, or comparison against an external motion-capture reference. The result establishes basic operability, not yet metrological motion-capture accuracy. Sixteen actuators, but no tactile thumb. The haptic layer consists of sixteen coin-shaped linear resonant actuators, each eight millimeters in diameter and 2.55 millimeters thick. The index, middle, ring, and little fingers each receive four actuators. They are positioned on the left, right, dorsal, and volar sides of the distal finger. The thumb is not included in this tactile array, and the slave has force–torque sensors on four fingertips rather than all five. That is a meaningful limitation for pinch and opposition tasks, where the thumb commonly supplies both a critical contact and a major fraction of the grasp wrench. STRP can capture thumb motion, but it does not close the same tactile loop around thumb contact. The actuators are attached using adhesive tape, structural adhesive, printed mounts, and small screws. An STM32F103 controller addresses the haptic drivers through an I²C architecture. The implementation is not elegant in the sense of custom flexible electronics, but it is accessible: ordinary printed parts, commodity vibration motors, magnets, fabric, and a relatively conventional microcontroller board. A sensorized dexterous slave. The slave is a commercial 17-degree-of-freedom ZWHAND mounted on a UR5. The hand weighs about 850 grams, has a rated firm-grasp payload of three kilograms, and communicates over RS485 using Modbus RTU. Motion commands, arm commands, and force measurements run at 100 hertz. Each of the four instrumented fingers carries a custom miniature six-axis force–torque sensor co-developed with the sensor manufacturer. The transducer is 9.5 millimeters in diameter and 17 millimeters tall. Its specified ranges are plus or minus 50 newtons for the two tangential axes, plus or minus 150 newtons axially, and up to one newton-meter on the largest torque channel. The quoted resolution is about 25 millinewtons tangentially and 124 millinewtons axially. A hemispherical cap and cylindrical sidewall surround the transducer. This geometry is important because it allows contacts from around the fingertip, rather than restricting measurement to a flat pad. The assembly effectively turns the distal finger into a compact force-sensing shell. The high measurement range provides overload margin, although it is much larger than the two-to-five-newton forces used in the cue-identification experiment and the roughly two-to-four-newton grip forces reported during beaker manipulation. The paper specifies less than three percent full-scale crosstalk, but dynamic crosstalk and drift during continuous sliding were not characterized. The authors explicitly identify that as future work. Turning a Force Vector into a Skin Pattern. At the beginning of an experimental session, STRP averages 20 unloaded readings to estimate the biases on the two lateral force channels. Those offsets are subtracted from subsequent measurements. A hysteresis gate turns feedback on at 0.2 newtons and keeps it active until force falls below 0.1 newtons. This prevents the motors from rapidly chattering near the detection threshold. The sign of each lateral component selects an actuator. Opposing signs drive opposing sides of the fingertip, while force magnitude controls vibration intensity. If two components are simultaneously present, two actuators run together. A dorsal-right contact therefore becomes a combined dorsal-and-right pattern rather than being quantized to the nearest cardinal direction. The amplitude mapping uses a compressive psychophysical curve derived from Stevens’ power law. Above the dead zone and below 20 newtons, the digital drive level increases with force raised to the power of 0.7. Above 20 newtons, it saturates. In practical terms, this expands perceptual differences among lower forces while compressing the upper end of the sensor range. The mechanical fixture aligns the sensor axes with the anatomical left–right and dorsal–volar directions. This eliminates a runtime coordinate transform, but it also clarifies what “spatial” means here. STRP reports contact in the robot fingertip’s local anatomical frame, not in the world frame. If the hand rotates, a “right” cue still means the right side of that finger. Inferring the world-space wall normal still requires the operator to combine tactile information with knowledge of the hand and wrist pose. There is also an important distinction between sensing dimensionality and rendering dimensionality. The transducers measure six-axis wrench data, but the online directional renderer primarily consumes the two tangential force channels. The torque channels are reserved for future contact localization and slip detection. The paper records the axial force channel during grasp experiments, but it does not clearly describe how that axial value participates in the four-actuator rendering rule. STRP should therefore be described as a six-axis-sensed, spatially vibrotactile interface, not as a system that renders all six wrench components. It does not reproduce torque, continuous three-dimensional force direction, distributed pressure, texture, or impedance. Its contribution is a compact four-quadrant representation of local contact direction and intensity. Latency and the Real Meaning of “Low Cost”. The tactile loop runs at 100 hertz. The authors estimate approximately five milliseconds for Modbus sensor readout, six milliseconds for host processing and USB transmission, eight milliseconds to refresh the four active actuator channels, less than one millisecond for haptic-driver register updates, and as much as another eight milliseconds of polling-phase delay. Together, those components produce the reported latency of roughly 28 milliseconds. That is a useful engineering result, but it is not a complete contact-to-perception measurement. The estimate excludes the LRA’s mechanical rise time, and the paper does not provide a latency distribution, tail latency, clock-synchronized measurement trace, or evaluation under network jitter. The setup is a local laboratory system connected through USB, Ethernet, I²C, and RS485—not a long-distance teleoperation link. The glove itself has a reported bill of materials of USD 110.95. Roughly USD 68 goes to the control board, with the remaining cost distributed across the Hall sensors, 16 LRAs, magnets, printed parts, resin components, fabric, adhesives, and screws. That compares favorably with the paper’s listed USD 600 cost for DOGlove and USD 5,999 for the SenseGlove Nova 2. But USD 111 is not the cost of the complete STRP platform. The ZWHAND is listed at approximately USD 6,950, the UR5 at USD 16,670, and the four force–torque sensors at about USD 830 each. With the host computer, the authors estimate a total near USD 30,000. The strongest cost claim is therefore an incremental one. For a laboratory that already owns an arm and dexterous hand, STRP offers a remarkably inexpensive master interface. Starting from zero, it is a mid-priced research platform whose custom slave sensing alone costs around thirty times the glove BOM. The comparison also omits fabrication labor, assembly time, fit customization, maintenance, and the cost of reproducing the custom sensor integration. What the Experiments Actually Show. All reported experiments were performed by one trained operator: a 25-year-old right-handed man with six hours of prior teleoperation experience. Thus, an experimental count such as (N=15) means 15 repeated trials by one person—not 15 independent operators. The authors appropriately present descriptive statistics and defer population-level inference to future work. Can a person decode the four directions?. For the primary perception test, the operator wore an eye mask and earphones. An experimenter applied two-to-five-newton contacts to the robot fingertip with a blunt probe, and the operator verbally identified the resulting cue. Before measurement, the operator received ten minutes of familiarization and continued until identifying six cues consecutively. Each cardinal direction was then presented 20 times in randomized order. Left and right were identified correctly in all 40 trials. Dorsal and volar each produced one error, giving 78 correct responses out of 80, or 97.5%. A supplementary test presented four diagonal patterns by activating pairs of motors. The operator correctly identified 17 of 20, or 85%. Most confusion occurred along the dorsal–volar axis, where the actuators have less physical separation than the lateral pair. This is convincing evidence that one trained wearer can learn the code quickly. It does not yet establish how performance varies across finger sizes, glove fits, novice users, extended sessions, simultaneous multi-finger contacts, or operators wearing protective equipment. Does feedback improve ordinary pick-and-place?. The next experiment used a 250-milliliter borosilicate beaker. The operator grasped it, lifted it, transferred it, and placed it in a target zone. Fifteen trials used multidirectional feedback and 15 used no haptic feedback, with condition order randomized. Completion time fell from 31.2 to 18.4 seconds, a 41% reduction. Peak grip force fell from 4.11 to 2.79 newtons, a 32% reduction, while mean grip force fell from 2.72 to 2.03 newtons. Slip events declined from 1.5 to 0.2 per trial, and success increased from 11 of 15 trials to 14 of 15. The authors interpret this as a shift away from conservative over-gripping toward more confident force modulation. These numbers are substantial, but the comparison is multidirectional feedback versus no feedback. It shows that the complete haptic condition helped this operator. It does not isolate whether direction was responsible, because there was no control condition that rendered the same contact magnitude without spatial information. Can directional touch support active exploration?. The wall experiment is the paper’s most distinctive demonstration. The robot explored an opaque cardboard enclosure measuring 0.6 by 0.8 meters at the base and 0.6 meters high. The operator could not see or hear the interaction. Three conditions were tested over 20 trials each. Multidirectional feedback used all four actuator locations. “Palm-Only” activated only the volar actuator. The third condition provided no haptics. With multidirectional feedback, the operator found both target walls in all 20 trials, taking an average of 23.6 seconds. With Palm-Only feedback, the left wall was detected 19 times, but the right wall only four times; the complete task succeeded in four of 20 trials. Without haptics, no trial succeeded. The multidirectional condition also enabled a repeatable search procedure—probe left, return to center, then probe right—without wrist reorientation. This experiment captures STRP’s core idea well: active touch depends on knowing which motion produced which contact. Directional cues let the operator build a small spatial model of an invisible enclosure. There is nevertheless an ablation caveat. Palm-Only does not preserve every contact while removing direction. It entirely suppresses dorsal and lateral contacts. The comparison therefore combines two effects: directional localization and broader sensing coverage. A stronger control would route the magnitude of every contact to one common actuator. The operator would always know that contact occurred but not where. That would isolate the value of spatial coding itself. Can the operator complete an occluded manipulation task?. The final task began with a water bottle already grasped. The operator moved it toward the right wall, used an extended little finger as a probe, and attempted to release the bottle onto a target plate. Each condition received ten trials with a 30-second limit. Multidirectional feedback achieved four successful placements out of ten. Palm-Only and no-feedback conditions achieved zero. In all multidirectional trials, the operator correctly found the right-wall boundary; the six failures happened during release, when the bottle slipped or tipped after detachment. That 40% success rate is more informative than a cleaner demonstration would have been. STRP solved the contact-localization problem but exposed the next bottleneck: controlled release. Vibrotactile direction can tell a person where the wall is, but it does not physically stabilize the object or automatically coordinate decreasing normal force across four fingers. Why This Matters for Embodied AI. For teleoperation, STRP can be viewed as a low-bandwidth contact-state estimator presented directly to the human nervous system. The raw sensor stream is high-dimensional, but the operator often needs only a few task variables: which finger is in contact, which side is loaded, whether load is increasing, and whether the contact is stable. That compression may be more useful than pursuing literal physical simulation. In an occluded search, the distinction between left and right can be more valuable than a highly accurate scalar force. In a fragile grasp, a clearly perceptible load trend may matter more than reproducing exact stiffness. The same logic applies to learning-from-demonstration systems. Better feedback can alter the distribution of collected actions: fewer collisions, fewer emergency corrections, lower force peaks, and clearer transitions between approach, contact, manipulation, and release. The policy does not necessarily need access to the operator’s haptic signal. It can benefit indirectly because the human generated better demonstrations. STRP also supplies potentially useful privileged labels. Fingertip wrench streams can identify contact phases, estimate overload risk, segment slip events, or supervise tactile representations. The unused torque channels could eventually support contact-point estimation and incipient-slip detection. The present paper does not train an imitation-learning or reinforcement-learning policy, however. Its results establish operator-side feasibility. Demonstrating value for robot learning would require comparing policies trained on otherwise matched demonstrations collected with and without the feedback loop. A Critical Reading of the Claims. The paper’s strongest supported claim is that four skin locations can convey local contact direction accurately enough for one trained operator to use during an occluded search. The wall experiment is simple, but it demonstrates a capability unavailable from vision in that setup. The broader claim that directional feedback cannot be replaced by magnitude-only vibration remains under-tested. The beaker experiment compares directional feedback against silence. The wall ablation makes non-volar contacts silent rather than collapsing all contacts into a shared magnitude channel. Both experiments favor the full system, but neither cleanly separates contact detection from contact localization. The single-operator design is the largest empirical limitation. Repeated trials can characterize consistency for that person, but they do not capture differences in tactile acuity, training rate, finger anatomy, hand dominance, strategy, fatigue, or risk tolerance. The 97.5% discrimination result should therefore be read as a prototype verification rather than a population estimate. Similarly, the system’s six-axis terminology requires care. Six-axis sensors are present, but two lateral channels drive the directional renderer, torque is unused, and the role of axial force in the haptic mapping is insufficiently documented. The interface is much closer to a two-dimensional contact-direction display than to a six-axis wrench display. The hazardous-environment framing is motivational rather than experimentally validated. The tasks use a beaker, a cardboard enclosure, and a water bottle. There are no energized components, corrosive chemicals, radiation constraints, thick protective gloves, communication dropouts, environmental sealing tests, or formal safety procedures. Local 28-millisecond performance also says little about operation over a delayed wide-area network. Durability remains open as well. Sessions were limited to roughly 30 minutes with rest periods, and no fatigue-related decline was observed during those sessions. Long-term sensory adaptation, adhesive reliability, printed-joint wear, calibration stability over thousands of cycles, and impact resistance were not tested. The authors propose stronger fiber-reinforced links, metal pins, broader user studies, and dynamic sensor characterization as future work. Finally, reproducibility is weaker than the low BOM might imply. The paper provides a supplementary demonstration video and states that supporting data and materials are contained in the article or supplement, but it does not advertise a public repository containing CAD, PCB files, firmware, ROS packages, or assembly instructions. The miniature force sensor is a custom co-development, and two authors are employed by the sensor company. Those relationships are disclosed, but an independent laboratory may still need vendor support to reproduce the complete fingertip assembly. Where STRP Sits Among Current Teleoperation Systems. Vision-based platforms such as AnyTeleop emphasize general hand-pose detection and retargeting across different robot hands. LEAP Hand similarly lowered the hardware barrier for dexterous-learning research, offering an approximately USD 2,000 open platform that can be used with vision or gloves. These systems prioritize accessible motion generation, but force sensing and operator-side spatial touch are not their central contributions. DOGlove, presented at RSS 2025, is a closer comparison. It combines 21-degree-of-freedom motion capture with cable-driven force feedback and fingertip vibration for under USD 600, and its hardware and software are open source. DOGlove can resist finger motion, giving it a kinesthetic capability absent from STRP. STRP instead distributes four cutaneous channels around each instrumented fingertip, making the contact side explicit without a force-transmission mechanism. CDF-Glove pushes the cable-driven approach to roughly USD 230. It reports 20 available finger states, approximately 200-millisecond feedback latency, validation across multiple robot hands, and downstream Diffusion Policy experiments. Its emphasis is improving teleoperated demonstration collection through force reflection rather than encoding within-fingertip contact direction. TAG, released as a 2026 preprint, moves in the opposite direction from STRP’s intentionally coarse code. It combines 21-degree-of-freedom magnetic tracking with a 32-actuator tactile array on each finger, allowing spatial patterns corresponding to contact geometry to be rendered at much higher resolution. ViHaTeleop, meanwhile, combines visual tracking, tactile-aware retargeting, and finger-wise LRA feedback in a USD 550 system oriented toward manipulation learning. STRP is therefore neither the only inexpensive haptic glove nor the highest-resolution tactile display. Its particular niche is the combination of real fingertip force–torque sensing, four-quadrant local direction cues, simple commodity LRAs, and a sub-30-millisecond local feedback path. Its design asks how little tactile hardware is required to recover useful spatial awareness. How I Would Extend the Platform. For researchers adapting this idea, five changes would substantially strengthen the system. Add the thumb and explicitly separate force channels. A complete hand should report thumb contact, normal load, lateral shear, and slip as distinguishable cues. Direction could remain spatial, while normal force might control a common intensity channel and high-frequency transients could indicate slip. Use the complete wrench locally before sending it to the operator. Force and torque together can estimate where the fingertip shell is loaded. The system could map that contact point continuously to the nearest tactile actuators rather than mapping only two force components. Pair human feedback with slave-side reflexes. The object-placement failures show that the human should retain strategic authority while the hand autonomously handles millisecond-scale grip stabilization, slip recovery, and gradual release. Haptic cues and local autonomy are complementary, not competing solutions. Run matched human-subject ablations. A proper study should include multiple novices and experts, a no-feedback condition, a scalar all-contact condition, spatial feedback, and ideally kinesthetic feedback. Contact detection rate, localization accuracy, workload, force exposure, collision count, learning rate, and retention should all be measured. Treat teleoperation as a data-generation pipeline. Log synchronized video, glove state, robot state, full fingertip wrench, rendered haptic commands, local reflex activations, and task outcomes. Policies trained from each feedback condition can then reveal whether better human awareness actually produces more learnable demonstrations. The Bottom Line. STRP is not yet a deployable system for electrical substations or nuclear decommissioning. Its evidence comes from one operator, controlled laboratory tasks, local communications, and prototype-grade printed hardware. The final occluded placement task succeeds only four times out of ten even with full feedback. What the paper does establish is more focused and arguably more valuable: contact direction can be compressed into a cheap, learnable tactile code, and that code can restore a form of active spatial reasoning when vision disappears. For embodied-AI researchers, the lesson is that haptic fidelity should not be measured only by how accurately an interface reproduces physical force. A lightweight display that preserves the right task variables may outperform a more elaborate interface that communicates the wrong ones. STRP’s four little vibration motors per finger are compelling not because they recreate touch, but because they preserve just enough of its geometry to change what the operator can do.