A growing body of work suggests that tactile sensing gives robot policies contact information that complements vision in dexterous manipulation. However, visuo-tactile robot data remains scarce: dexterous demonstrations require teleoperating robots, which limits dataset scale. Human demonstrations are far cheaper to collect and offer a path to scale this data, but only if human and robot hands carry tactile sensors with corresponding signals. This requires sensors that conform to different hand geometries, cover the full hand, and share a common layout across embodiments. We present SoTa, a low-cost capacitive tactile skin that provides full-hand coverage on humans and robots while preserving a shared layout of 202 taxels across corresponding finger and palm regions. Our multilayer design with fabric electrodes enables in-house fabrication of thin, soft skins with customizable geometry for under $10 in materials per skin. The sensor retains over 97% of its initial response span after 10,000 loading-unloading cycles with traces retaining continuity through 1,280 tight-fist folding cycles. The shared taxel layout supports human-robot co-training with a common tactile encoder and no learned cross-sensor mapping. Across three contact-rich manipulation tasks, tactile observations improve in-distribution success over vision-only policies. With a fixed robot demonstration budget, adding human demonstrations more than doubles mean success across eight evaluation conditions, from 22.8% to 45.9%, improving success in all five out-of-distribution conditions. We plan to open-source the resources needed to fabricate and operate these skins.
Figures & tables
System
Sensor
Coverage
Sensing
Embod.
Aligned
Custom
elements (total)
H/R
fit
Robot-mounted tactile skins (no human-side wearable)
ReSkin, AnySkin [ 20 , 21 ]
Magnetic
Patch
5/patch
R
–
no (one-size)
AllTheFeels [ 22 ]
Magnetic
FT+P
56
R
–
no (one-size)
DexSkin [ 23 ]
Capacitive
Round Finger
120
R
–
yes (per-gripper)
Sharma [ 24 ]
uSkin (Magnetic)
FH
368
R
–
no (one-size)
TABLE I : Comparison of tactile interfaces relevant to SoTa. The upper block lists robot-mounted tactile skins (without human-side wearables). The lower block lists wearable interfaces with a human side (human-only or bridging human and robot embodiments). In each block, prior systems are chronologically ordered. Coverage : FT = fingertip, FT+P = fingertips and palm, FH = full hand (all phalanges + palm). Embodiment : H = worn by human, R = mounted on robot, H+R = both. Aligned H/R : HW = hardware/layout aligned, SW = learned alignment, – = not addressed. Custom fit : per-hand physical customization vs. one-size-fits-all.
Fig. 2 : SoTa tactile skin hardware. (a) Robot (left) and human (right) skins with custom-fit form factors and matched region-level taxel layouts. (b) Exploded view showing stack-up of the multilayer capacitive sensor. (c) Sensor responses over three consecutive cycles per load from 20 mN to 10 N; boxed inset shows the 20 mN detail. (d) Sensor capacitance over 10,000 loading-unloading cycles with a peak force of 8 N. The dense band’s upper and lower edges represent loaded and unloaded sensor outputs; the approximately rectangular band indicates stable cyclic response.
Fig. 3 : Our data collection setup. (a) Robot teleoperation. A Meta Quest controller tracks the operator’s wrist pose and a Rokoko smart glove tracks per-finger joints; both streams are fused at 30 Hz and command the dexterous Kinova + Wuji hand setup. A front-facing ZED Mini camera and our tactile skins record synchronized observations. (b) Human demonstration collection. The operator wears a Rokoko glove and our human-fit tactile skin while performing the same tasks without the robot. Human hand configurations were recorded for an exploratory hand pose prediction objective but are not used in our method or the reported experiments.
Fig. 4 : Policy architecture built on FTP-1. Image, prompt, and tactile observations condition the action expert, which generates actions through flow matching. A training-only language expert predicts task-phase labels from visual-language and tactile features. The action expert does not attend to the language expert, so the auxiliary branch is omitted at inference.
Fig. 5 : Objects used in our experiments. (a) In-distribution objects. We collected both robot and human demonstrations using these objects. (b) Objects seen in human demonstrations, but not robot demonstrations. (c) Objects used in OOD evaluation, which includes held-out objects and configurations.
Method
Policy reads tactile
Two-stage training
Human vision data
Human tactile data
RV
✗
✗
✗
✗
RV,T
✓
✗
✗
✗
RV,T+LV,T
✓
✓
✗
✗
RV,T+HV
✓
✓
✓
✗
RV,T+HV,T
✓
✓
✓
✓
TABLE II : Training configurations in different methods. All methods use robot demonstrations for action supervision. Two-stage training consists of representation pretraining, followed by joint action and auxiliary language learning. Human data contribute only to the auxiliary objective. Green ticks ( ✓ ) indicate inclusion, red crosses ( ✗ ) indicate absence.
Fig. 6 : Effect of tactile feedback. We present the success rates achieved by policies trained on robot data only with and without tactile sensing. We test our policies on three tasks with in-distribution and out-of-distribution objects. The y-axis shows the percentage of fully successful rollouts out of 40 trials. Partial completions count as failures. Results show that tactile feedback improves dexterous policy learning for in-distribution settings, while its effects in out-of-distribution settings are mixed.
Fig. 7 : Auxiliary supervision and human-robot co-training. We present quantitative results supporting that learning dexterous visuo-tactile manipulation policies via human-robot co-training ( RV,T+HV,T ) outperforms policies trained without human data ( RV,T ) in every reported condition, with gains concentrated in OOD settings. We also show ablation results on variants of our full method. RV,T+LV,T denotes two-stage robot-only training with visuo-tactile language supervision. RV,T+HV denotes an ablation that does not utilize tactile inputs in the auxiliary objective.
Fig. 8 : Conductor durability during repeated tight-fist folding on a Wuji hand. Curves show mean resistance of four conductors. Dashed arrows indicate the first open-circuit reading ( >40MΩ ) in any trace, observed at 80 cycles for copper-polyimide and 640 cycles for silver on TPU. SoTa’s electrode and jumper wires reach the final checkpoint at 1,280 cycles with finite resistance.
Finger
RMS of ΔC/C0 (%)
flexed
Flexed finger
Unflexed fingers
Palm
Thumb
3.66±0.63
1.69±0.89
2.30±0.87
Index
6.18±1.10
1.54±0.35
1.60±0.46
Middle
5.49±0.47
2.55±0.48
1.20±0.20
Ring
5.60±0.56
1.97±0.62
1.34±0.33
Little
4.09±0.86
1.45±0.41
1.60±0.44
TABLE III : Motion-induced responses during individual finger flexion. Values are regional RMS of ΔC/C0 (%) over ten flex–relax cycles, reported as mean ± SD across four experimenters. Each 60s trial comprised a 10s still-hand baseline followed by ten cycles of 2s flexion and 3s relaxation; RMS was computed over the 50s after the baseline. The unflexed-finger region excludes the flexed finger and palm.
Dexterous manipulation requires tactile feedback. However, robot tactile demonstrations are difficult to scale,because dexterous-hand teleoperation provides limited tactile feedback to the operator. In contrast, human demonstrations offer a substantially more scalable source of diverse tactile interactions. Motivated by a simple premise: hands can change, but the underlying physics of interaction does not. We leverage human tactile data to improve dexterous manipulation policies. Specifically, we first build a tactile motion-capture system that synchronously records images, tactile signals, and hand motions. Using this system, we construct the UVTA dataset spanning five contact-rich tasks, with 1,000 human demonstrations covering diverse interaction patterns and 150 robot demonstrations per task. To transfer the underlying physics of human interaction to robot control, we propose a Unified Visual-Tactile-Action Model that maps both embodiments into aligned tactile and action representations and jointly predicts future action and tactile trajectories. The joint objective enables human demonstrations to supervise contact-aware representation learning, while only robot actions are executed during deployment. In real-robot evaluations across five tasks, our method achieves an average success rate of 70%, outperforming the strongest visual-tactile baseline, which achieves 29%, and an architecture ablation, which achieves 42%. Performance improves consistently with additional human demonstrations and exhibits no saturation at 1,000 demonstrations per task, validating the effectiveness of scalable human tactile data for dexterous manipulation. Project page is available at https://uni-vta.github.io/.
Wenqiao Li, Qianyou Zhao, Jiawen Hao +5
Shanghai Jiao Tong University · Sharpa Robotics · Beijing Institute of Technology +1
Collecting tactile demonstrations on robots is costly and slow, motivating the use of lower-cost human tactile gloves for scalable data collection. However, human capacitive/piezoresistive gloves and robotic tactile sensors differ fundamentally in transduction principle, sensor layout, spatial resolution, and dynamic response, making alignment of raw sensor channels ill-posed. To address this problem, we present TACROSS, a scalable system for learning from human touch and transferring it to robots that bridges this heterogeneity by aligning tactile streams at the level of contact events rather than raw sensor values. The hardware component of TACROSS integrates a piezoresistive glove with five layers and a cost of USD 10.86 with 285 sensing points. To align contact semantics, we design canonicalizers and residual adapters that map heterogeneous signals into a shared tactile latent with 256 dimensions via a temporal Transformer with attention across fingers. We further introduce a robot-grounded policy learning scheme in which robot demonstrations provide the sole source of ground-truth action supervision, while human demonstrations support tactile representation learning and provide confidence-weighted auxiliary supervision through valid retargeted hand targets. We evaluate our system on four contact-rich manipulation tasks. Compared to conventional teleoperation, our proposed system achieves a 3.5-fold efficiency improvement while reducing demonstration acquisition equipment cost by 95.7%. We will open-source the TACROSS hardware and software system and publicly release a tactile dataset comprising over 150 hours of recordings. Project page: https://tacross-touch-project.github.io/.
Bo Chen, Huanzhang Hu, Junyang Ma +11
The University of Hong Kong · INFIFORCE · The Chinese University of Hong Kong (Shenzhen) +3
Vision-based learning from demonstrations has achieved remarkable success in enabling robots to perform manipulation tasks and high-level semantic reasoning, yet it remains insufficient for complex, contact-rich manipulation. While there is broad agreement that tactile sensing improves manipulation, there is no empirical guidance on which tactile sensors are best suited for which manipulation tasks. In this paper, we provide a systematic, task-driven evaluation of tactile sensors for robot manipulation and propose a framework for selecting and evaluating sensors based on manipulation policy performance. Separate manipulation policies are trained for tactile sensors of four distinct modalities: visual, acoustic, magnetic, and resistive, across three tasks: pick-and-place with unknown mass, object reorientation, and plug insertion. For each task, an analysis of how sensor properties such as spatial resolution, shear sensing, and tactile representation, and the inherent material friction affect task performances is done. Rather than tactile sensing being universally beneficial in the same way, our results show that the usefulness of tactile information depends strongly on sensor modality, material properties, and the specific manipulation tasks. All of the tactile sensors, code, data, and hardware setup will be publicly available on the project website.
Anya Zorin, Zilin Si, Myungsun Park +11
UC San Diego United States · CMU · UC San Diego +2