Pranav Ponnivalavan, Satoshi Funabashi,
Alexander Schmitz, Tetsuya Ogata, Shigeki Sugano
Waseda University
2026 IEEE International Conference on Robotics & Automation (ICRA), 1–5 June, Vienna, Austria
Abstract
Humans can achieve diverse in-hand manipulations, such as object pinching and tool use, which often involve simultaneous contact between the object and multiple fingers. This is still an open issue for robotic hands because such dexterous manipulation requires distinguishing between tactile sensations generated by their self-contact and those arising from external contact. Otherwise, object/robot breakage happens due to contacts/collisions. Indeed, most approaches ignore self- contact altogether, by constraining motion to avoid/ignore self- tactile information during contact. While this reduces complexity, it also limits generalization to real-world scenarios where self-contact is inevitable. Humans overcome this challenge through self-touch perception, using predictive mechanisms that anticipate the tactile consequences of their own motion, through a principle called sensory attenuation, where the nervous system differentiates predictable self-touch signals, allowing novel object stimuli to stand out as relevant. Deriving from this, we introduce TaSA, a two-phased deep predictive learning framework. In the first phase, TaSA explicitly learns self-touch dynamics, modeling how a robot’s own actions generate tactile feedback. In the second phase, this learned model is incorporated into the motion learning phase, to emphasize object contact signals during manipulation. We evaluate TaSA on a set of insertion tasks, which demand fine tactile discrimination: inserting a pencil lead into a mechanical pencil, inserting coins into a slot, and fixing a paper clip onto a sheet of paper, with various orientations, positions, and sizes. Across all tasks, policies trained with TaSA achieve significantly higher success rates than baseline methods, demonstrating that structured tactile perception with self-touch based on sensory attenuation is critical for dexterous robotic manipulation.
Paper
arXiv (PDF): https://www.arxiv.org/pdf/2602.05468
IEEE Xplore: (add link after publication indexing)
Accepted in ICRA 2026 as a conference paper
Bibtex
@article{tasa2026,
title = {TaSA: Two-Phased Deep Predictive Learning of Tactile Sensory Attenuation for Improving In-
Grasp Manipulation},
journal = {arXiv preprint arXiv:2602.05468},
year = {2026},
url = {https://arxiv.org/abs/2602.05468}
}
Overview Video
Method
The framework has two stages. First is Self-Touch Learning, where a fully connected network (FCN) learns to predict tactile signals caused only by the hand touching itself (finger–finger or finger–palm). The FCN takes current joint positions and commanded/desired joint positions (8 active joints total, input size 16), and outputs predicted tactile signals for index and thumb tips plus an auxiliary joint-state output. Since this phase uses only free-space/self-contact data (no objects), all tactile readings are treated as self-touch targets. By training across open-pinch, closed-pinch, and rubbing motions, the FCN learns how joint configurations map to characteristic self-contact tactile patterns.
Second is Motion Learning, where an LSTM uses three inputs at each timestep: current joints, raw tactile sensing, and the FCN’s predicted self-touch. This helps separate self-generated tactile effects from object-induced contact, giving cleaner signals for manipulation. The LSTM predicts next joint state, desired joint targets, and next raw tactile output; then these predicted postures are fed into a frozen self-touch FCN to generate next-step self-touch as a differentiable forward model. Including commanded joints and an auxiliary joint-state consistency signal improves kinematic plausibility and prediction quality. The resulting method (self-touch + raw tactile, st-rnn) is compared against a raw-tactile-only baseline (t-rnn).
Key idea: By explicitly predicting expected self-touch, the policy can focus on unexpected/object-induced tactile events, reducing confusion in contact-rich manipulation.
Results & Analysis
Across all three tasks, adding self-touch modeling to raw tactile sensing (TaSA: RT+Self) gives a clear and consistent performance boost over the RT-only baseline. In Paper Clip Fixing, success rises from 70% (42/60) to 95% (57/60), which is a large gain in a relatively small trial set where each failure is costly. In Coin Insertion, performance improves from 68% (61/90) to 92% (83/90), showing that the benefit is not task-specific but transfers to a different contact pattern and geometry. Together, these two tasks suggest that self-touch compensation helps the policy focus on external contact cues rather than being confused by tactile signals produced by the hand’s own motion.
The strongest evidence appears in Pencil Lead Insertion, the hardest and most precision-sensitive scenario: success jumps from 26% (66/250) to 58% (146/250). Although this task remains challenging, the improvement is substantial and indicates that TaSA is especially valuable when fine contact interpretation matters most. In practical terms, RT+Self improves robustness by reducing ambiguity in tactile perception, thus, separating internally generated touch from object interaction, so the controller can make better temporal decisions during manipulation. This pattern of gains across easy-to-hard tasks supports the claim that self-touch-aware tactile learning is a key factor behind more reliable contact-rich manipulation.
Additional observations:
• Self-touch prediction tracks tactile signals closely during sustained self-contact, with low error except around rapid transitions.
• PCA feature-space plots show clearer class separation when self-touch is modeled, indicating reduced nuisance variance from self-generated contact.
• Generalization improves on unseen test splits (e.g., perfect middle-slot success in coin insertion under RT+Self).
Paper Clip Fixing (Fig. 1(a) and 1(b)): In Case A (RT), the initial tactile features are strongly influenced by grasp preload and finger–finger shear. As a result, trials from the back and front placements form clusters that are somewhat distinct but still overlap, indicating that self-touch effects obscure placement-specific differences. In Case B, incorporating the self-touch prediction reduces the influence of self-generated signals, which lowers variance along the axis aligned with finger–finger contacts and increases the inter-centroid distance between front and back for each size. Mechanically, the front placement induces finger–finger contact that mimics clip–paper completion; incorporating self-touch mitigates this, yielding cleaner boundaries. Coin Insertion (Fig. 2(a) and 2(b)): Although the pre-insertion state is similar across the train and test slots, adding the self-touch prediction yields a stabilized embedding: clusters translate coherently and several classes tighten (notably 1 yen and 500 yen right), indicating reduced nuisance variance from finger–finger preload. The 100 yen and 500 yen groups still partially merge—their thicknesses differ by only ≈0.1 mm, so RT only produces near-identical normal forces and contact-patch areas—but in Case B the covariance shrinks and spacing increases, making slot boundaries less entangled. This representation already supports better alignment and generalization: RT+Self attains improved results on the unseen middle slot across all coin sizes. Lead Insertion (Fig. 3(a) and 3(b)): In CaseA, the diameter pattern is already visible, but the 0.7 mm and 1.3 mm groups partly overlap. Their small contact areas and weak normal forces give low signal quality, and even slight angle changes (±20◦) alter the shear direction enough to blur diameter differences. Case B keeps the diameter separation clearer and also shows some angle-based grouping, but the thinnest leads are still the hardest to tell apart. This explains why success is lower at larger angles and with very thin leads.
Acknowledgements
This research was supported by Moonshot R&D with a grant number of JPMJMS2031. All authors are with Waseda University, Okubo 3-4-1, Shinjuku, Tokyo 169-8555, Japan (pranavponni@fuji.waseda.jp).
