Design of a Biomimetic Joint-Covering Skin with Tissue-Like Structure to Enhance Proprioception in a Musculoskeletal Humanoid

• Accurate proprioception is essential for stable behavior and adaptive motion in musculoskeletal humanoids.Source: §I · p. 1
• Joint-angle estimation is difficult in musculoskeletal humanoids because their complex structures prevent straightforward direct measurement.Source: §I · p. 1
• Existing approaches often rely on limited sensor numbers and modalities, including muscle-length and tension sensors.Source: §I · p. 2
• Biological proprioception integrates signals from mechanoreceptors in muscles, skin, and surrounding joints, providing redundancy and supporting discrimination of external disturbances.Source: §I · p. 2
• The study designs and evaluates layered joint-covering skin with pressure- and stretch-sensitive elements on Musashi-W, both independently and alongside muscle sensing.Source: §I · p. 2

• Biological skin distributes multiple mechanoreceptor types across layered tissue to integrate diverse tactile and proprioceptive information.Source: A. Biological Basis of Skin… · p. 2
• Existing biomimetic tactile sensors primarily target object recognition and contact-state estimation rather than proprioception.Source: B. Engineering Approaches to Cutaneous… · p. 2
• Whole-body modular skins primarily support contact detection and tactile mapping, while proprioception is not their main focus.Source: B. Engineering Approaches to Cutaneous… · p. 2
• Flexible e-skins provide large-scale sensing but face implementation complexity and possible effects on joint mechanics when adapted to three-dimensional joints.Source: B. Engineering Approaches to Cutaneous… · p. 2
• Embedded sensors in soft bodies are typically arranged to achieve specific functions with a minimal number of informative sensors.Source: B. Engineering Approaches to Cutaneous… · p. 2
• The proposed design follows biological organization by embedding numerous receptor-like elements within supporting tissue rather than optimizing a grid-like sensor layout.Source: B. Engineering Approaches to Cutaneous… · p. 2

• The design embeds numerous receptors within a layered tissue structure that mechanically supports them.Source: §III · p. 3
• Fabrication embeds sensors during multilayer construction, combines paired skin components, coats the epidermis, encloses the joint, and mounts the structure on Musashi-W.Source: Figure 2 · p. 3
• The skin mimics epidermis, dermis, and subcutaneous tissue using a multilayer soft structure.Source: A. Tissue-like Layered Structure and… · p. 3
• Material selection targets the biological tendency for compliance to increase from the outer layer toward inner layers.Source: A. Tissue-like Layered Structure and… · p. 3
• The implementation uses silicone rubber for the epidermis, silicone gel for the dermis, and foamed silicone sponge for subcutaneous tissue.Source: A. Tissue-like Layered Structure and… · p. 3

• The study selects Ruffini-like stretch sensors and Merkel cell-like elements because spatial constraints prevent reproducing all biological cutaneous receptors.Source: B. Receptor-like Elements · p. 3
• Strain gauges serve as Merkel cell-like elements by changing electrical resistance under applied strain.Source: B. Receptor-like Elements · p. 3
• A conductive filament chainmail structure provides a stretch-responsive sensor corresponding to Ruffini endings.Source: B. Receptor-like Elements · pp. 3–4
• Merkel-like strain gauges are embedded between epidermal and dermal layers, while Ruffini-like structures are placed within the subcutaneous layer.Source: B. Receptor-like Elements · p. 4

• Fiber bundles embedded in the subcutaneous layer mimic collagen structures and provide mechanical continuity with the humanoid’s skeletal structure.Source: C. Development of Tissue-Structured… · p. 4
• Thirty-two strain-gauge sensors were positioned at the epidermis-dermis interface in a semi-random distribution modeled on biological receptor organization.Source: C. Development of Tissue-Structured… · p. 4
• Twelve conductive chainmail structures representing Ruffini-like elements were embedded in the subcutaneous layer and oriented along the fiber bundles.Source: C. Development of Tissue-Structured… · p. 4
• Dragon Skin 30 was coated and cured twice to form the epidermal layers, producing a pair of biomimetic skin components.Source: C. Development of Tissue-Structured… · p. 4
• The skin components enveloped an open-structure biomimetic joint and were connected at their interface using Soma Foama 15.Source: C. Development of Tissue-Structured… · p. 4
• Musashi-W integrates joint-covering skin signals with muscle-derived length and tension measurements as proprioceptive information.Source: C. Development of Tissue-Structured… · p. 4

• Sensorimotor data were collected while operating Musashi-W equipped with the tissue-structured biomimetic skin.Source: A. Sensorimotor Data Collection · p. 4
• The right upper limb contains five muscle-module pairs actuating the shoulder and elbow, providing 10 muscle-length and 10 muscle-tension sensors.Source: A. Sensorimotor Data Collection · p. 4
• The skin contributed 32 Merkel-cell-like strain-gauge signals and 12 Ruffini-like stretch-sensor signals for estimating elbow joint angles.Source: A. Sensorimotor Data Collection · p. 4
• Motion capture replaced embedded potentiometers and measured elbow pitch, yaw, and roll angles using markers on the upper arm and forearm.Source: A. Sensorimotor Data Collection · p. 4
• Body-schema-driven motion produced varied postures despite the modified body structure, while the tissue-biomimetic joint enabled elbow roll motion.Source: A. Sensorimotor Data Collection · pp. 4–5
• Approximately one hour of operation produced 615 sensorimotor samples covering about 50° of elbow motion along each axis.Source: A. Sensorimotor Data Collection · p. 5
• Muscle-related, cutaneous, and motion-capture measurements were recorded simultaneously to construct the proprioceptive-estimation dataset.Source: Figure 3 · p. 5

• Three-degree-of-freedom elbow angles were estimated separately from muscle length, muscle tension, Ruffini-like skin, and Merkel-like skin signals.Source: B. Single-Modality Proprioceptive… · p. 5
• The dataset was split into 70% training, 15% validation, and 15% test sets across 20 random seeds.Source: B. Single-Modality Proprioceptive… · p. 5
• A fully connected MLP with three 256-unit hidden layers predicted the three joint angles using ReLU activations, dropout, and a linear output layer.Source: B. Single-Modality Proprioceptive… · p. 5
• RMSE was computed for each joint-angle dimension and averaged across dimensions, with distributions shown by modality and summary statistics reported across seeds.Source: B. Single-Modality Proprioceptive… · p. 5
• The Wilcoxon signed-rank test with Holm correction compared paired RMSE values across modalities because normality could not be reliably assumed.Source: B. Single-Modality Proprioceptive… · p. 5
• Muscle length had the smallest mean and median RMSE, significantly outperforming the other modalities with Holm-corrected p < 0.001.Source: B. Single-Modality Proprioceptive… · p. 5
• Approximately 3°: skin modalities generally produced RMSE values within a range of about three degrees, though larger than muscle length.Source: B. Single-Modality Proprioceptive… · p. 5

• The experiments defined muscle-only, skin-only, and combined muscle-plus-skin modality conditions.Source: C. Multimodal Proprioceptive Estimation · pp. 5–6
• Concatenation fusion directly joined modality features before a shared MLP, whereas encoder fusion projected each modality into a 32-dimensional embedding before regression.Source: C. Multimodal Proprioceptive Estimation · p. 6
• Multimodal performance used RMSE from 20 random seeds and compared each condition against muscle length as the baseline with paired statistical testing.Source: C. Multimodal Proprioceptive Estimation · p. 6
• M+T enc achieved higher estimation accuracy than muscle-length-only estimation, with Holm-corrected p < 0.001.Source: C. Multimodal Proprioceptive Estimation · p. 6
• Other modality-fusion conditions did not consistently outperform the muscle-length-only condition.Source: C. Multimodal Proprioceptive Estimation · p. 6

External stimulation

• External stimulation was applied to the elbow with a stick or by hand during robot motion, producing disturbance data for evaluation.The procedures used approximately 15 minutes of sensorimotor data collection under pressure from various directions.Source: D. Evaluation of Joint-Covering Skin… · p. 6

Evaluation procedure

• Models trained without disturbances were evaluated on sensor data collected under external mechanical stimulation.Single-modality training and estimation procedures matched those used in the preceding experiments.Source: D. Evaluation of Joint-Covering Skin… · p. 6

Results

• Estimation errors increased for both skin and muscle modalities under external disturbances, with particularly large degradation for skin.External contact introduced deformation that differed from the conditions used to train the models.Source: D. Evaluation of Joint-Covering Skin… · p. 6
• The muscle modality retained a mean error of approximately four degrees under disturbance conditions.This indicates that external mechanical stimuli had a relatively limited influence on muscle signals in this experiment.Source: D. Evaluation of Joint-Covering Skin… · p. 6

Evaluation metric

• RMSE values in the external-stimulation comparison were averaged across the three elbow joint angles and paired by random seed over 20 seeds.The comparison included muscle length, Merkel-like skin, and Ruffini-like skin sensing modalities.Source: Figure 6 · p. 7

Proprioceptive information

• Skin-only estimation achieved joint-angle accuracy of approximately three degrees, indicating that deformation around the joint contains proprioceptive information.The result is consistent with biological evidence that skin stretch sensed by cutaneous mechanoreceptors contributes to proprioception.Source: §V · p. 7

Multimodal integration

• Encoder-based fusion of muscle and skin modalities outperformed the muscle-length-only condition, suggesting complementary sensory information.The improvement depended on the fusion architecture; other fusion conditions did not consistently outperform muscle-length-only sensing.Source: §V · p. 7

Disturbance interpretation

• External contact strongly affected skin estimates because local contact deformation was superimposed on deformation caused by joint motion.This shifted the input distribution relative to models trained under non-stimulated conditions.Source: §V · p. 7
• Comparing skin- and muscle-derived estimates may help indicate external contact or muscle-side abnormalities when the modalities diverge.The proposed interpretation relies on muscle estimates remaining relatively stable in one example and skin estimates remaining stable in the other.Source: §V · p. 7

Broader implications

• The joint-covering skin may provide mechanical protection, contact detection, and complementary sensing for robots interacting physically with their environments.The proposed utility extends beyond musculoskeletal humanoids to general robotic systems.Source: §V · p. 7

Limitations

• The implementation is a conceptual reconstruction rather than a strict biological reproduction because its sensors are larger and less numerous than biological mechanoreceptors.Future work should improve sensor miniaturization and density.Source: §V · p. 7
• Long-term stability and high-speed-motion response were not fully evaluated, while hysteresis, fatigue, drift, and quasi-static data constrain current validation.Dynamic-motion evaluation, online calibration, adaptive learning, and time-series models remain future work.Source: §V · p. 7

Contribution

• The study presented a biomimetic joint-covering skin with a tissue-like structure aimed at enhancing proprioception in a musculoskeletal humanoid.The design integrates numerous receptor-like elements into soft tissue rather than relying on a few high-performance sensors.Source: §VI · pp. 7–8

Findings

• Skin-derived information improved joint-angle estimation when integrated with muscle sensing.The experiments also demonstrated meaningful estimation using receptor-like elements embedded within the tissue structure alone.Source: §VI · p. 8

Implications

• The flexible skin covering may mitigate direct transmission of external stimuli to muscles and contribute to interpreting those stimuli.The conclusion frames these functions as possible contributions of the joint-covering structure.Source: §VI · p. 8