Peripheral control enabled by distributed sensing in an octopus-inspired soft robotic arm for autonomous underwater grasping

Abstract

Octopuses exhibit remarkable motor dexterity through distributed sensing and control in their flexible arms. Inspired by this biological model, we present a tendon-driven soft robotic arm featuring optoelectronic mechanosensors embedded in suction cups and a hierarchical behaviour-based control architecture. Each artificial sucker integrates light-emitting diodes and phototransistors to detect contact force and direction via light reflection, achieving high sensitivity (~400 mV N⁻¹ in the 0–2-N range) and directional accuracy (error, <18°). The sensors operate reliably in dry and wet environments with minimal drift and hysteresis. Their compact, modular design facilitates system integration. The hierarchical control architecture enables local reflexes at the suction cup level and global coordination for autonomous grasping. The system reliably detects contact events, estimates force and direction and infers object position relative to the arm, enabling purposeful interaction. This work advances sensor-integrated soft robotics, demonstrating the potential of biologically inspired designs for adaptive grasping in unstructured environments.

Main

A central goal of robotics is to endow artificial physical systems with sufficient autonomy and functionality to operate safely and effectively in real-world environments. However, the challenges posed by unstructured, dynamic and unpredictable scenarios, such as those in urban or natural conditions, continue to limit the adaptability and deployment of robots outside controlled industrial contexts. By contrast, biological systems provide interesting models of robust and plastic behaviour in such environments. A critical observation of natural beings highlights the role of soft, compliant structures in shaping behaviours and processing information through close interaction with the body and the environment^1,2,3. The survival of any individual relies on distributed perception, including tactile sensing, which plays a fundamental role in enabling close-loop control for both locomotion and manipulation¹. The capacity of a physical system to gather and interpret sensory information from its surroundings is essential for adaptive and intelligent behaviour⁴.

A paradigmatic example of adaptive coevolution of the nervous system and body morphology is provided by the octopus. The octopus has a fully hydrostat body and eight muscular continuum arms with infinite degrees of freedom, whose complexity is simplified through the control of basic motion primitives such as elongation, shortening, bending and torsion^5,6,7. Each arm is equipped on its ventral side with one or two rows of suckers—depending on the species—capable of generating strong attachment forces on non-porous surfaces⁸. The suckers have a dense sensory innervation⁹ and rich peripheral nerve branching from the axial nerve cord to the sucker ganglion (SG)¹⁰. The SG has been hypothesized to be a reflex control centre for sucker motor programs, but it appears that neighbouring-sucker coordination must also involve the axial nerve cord¹¹. It has been suggested that much of the sensory information is processed in the peripheral nervous system (PNS) of the arm, architected into several local neural circuitries (LNC) from the suckers, and activating stereotypical movements without information necessarily passing through the central nervous system (CNS), in contrast to skeletal animals¹². A special division of labour between the CNS and PNS of the arms has been demonstrated^13,14,15, although CNS seems to control planning and execution of goal-directed movements, such as fetching motions¹⁶. The combination of high dexterity, distributed sensing and peripheral information processing makes the octopus an appealing and inspirational model for the development of autonomous robotic manipulators capable of rich configuration capabilities and decentralized, sensor-driven control architectures.

The key principles of embodied organization revealed through the study of octopus motor control have inspired soft robotics research for several years^17,18,19,20. Examples span from soft manipulators and grippers acting in air, wet and dirty conditions^{19,21,22,23,24,25,26,27} to locomoting systems^24,28,29. Local reflex circuitry in octopus suckers has been recently imitated through pneumatic logic gates and pressure feedback in fluidically actuated fingers³⁰. However, integrating octopus-like dexterity and distributed tactile sensing via suction cups in soft robotic arms remains an open challenge. Current octopus-inspired robotic arms with suction cups face a trade-off between integrating distributed sensing and preserving dexterity, for example, by limiting arm manoeuvres to a single bending plane or restricting sensing functionalities to simple attachment/detachment detection²², without measuring contact force or direction. This integration poses challenges in appropriately dimensioning the electronics without compromising the compliance and manoeuvrability of the soft body, as well as keeping fabrication complexity low.

Among various sensing approaches³¹, optical sensors appear particularly suited for integration in soft bodies. They measure light intensity and typically consist of a light source, a photodetector and a waveguide that optically couple the two components. When embedded in a soft, reflective body, arrays of light-emitting diodes (LEDs) and photodetectors can detect contact events by sensing local deformations that alter light reflection and signal intensity³². Similar sensing can be achieved using optical fibres, either to transmit light between a source and camera through a deformable reflective surface or directly as force sensors that modulate light intensity when deformed^33,34,35,36. On the basis of this principle, optoelectronic soft tactile skins can be realized by utilizing an elastomeric layer as both mechanical substrate and waveguide, integrating LEDs and photodetectors along the perimeter^37,38.

These strategies illustrate the potential of optical methods for tactile sensing in soft systems, yet they leave open how such sensing can be exploited to drive rich, adaptive behaviours. To address this open challenge, this paper presents a tendon-driven soft robotic arm that uniquely couples distributed optical sensing with adaptive behaviour generation, preserving compliance and enabling rich, multimodal interaction. Our system balances sensing accuracy, compactness and seamless mechanical integration, allowing the arm to distinguish and utilize diverse sensory information (contact/no-contact, force magnitude, direction) in sensory-motor activation of complex movements (ventral/dorsal bending, twisting). The integrated sensing elements and electronics provide reliable, reproducible signals that support a bioinspired hierarchical control system, enabling tactile-enriched, adaptive and intelligent behaviour (Fig. 1).

Fig. 1: Biological versus artificial model.

Juxtaposition between the biological model and its relevant features and the artificial system. a, The octopus’s dexterity is replicated in the artificial soft arm to perform different grasping configurations (ventral and dorsal bending and twisting). The artificial arm imitates the distributed but hierarchically organized control architecture of the octopus. The octopus has a central, small brain that orchestrates high-level behaviour within the CNS, to which the PNS connects via an inter-arm connection (IAC) circuit for multiarm coordination. b, PNS is distributed along the arms and architected into the LNC located in the suckers. It contains about 350 million cells, comprising about two-thirds of all neurons in the octopus. Most of the cells are located in the axial nerve cords projecting from the brain to the arms⁶¹. c, In the robot, artificial suction cups are scaled to facilitate integration in a conical arm, similar to the natural one. Mechanoreceptors are distributed across suction cups to detect contact, its direction and the force on each cup, providing feedback for autonomous grasping. Local processing triggers local reflex-like behaviour (that is, suction) as signals are passed to peripheral control; these signals are combined to define the tendon actuation strategy. d, Some instances in the arm configuration space are shown in successive sequences in the bottom-right corner.

The robotic arm mimics the morphology of an octopus arm, with a conical shape (410 mm in length and 40 mm in diameter at the base), and ten sensorized suction cups decreasing in size from base to tip (from 20 mm to 12 mm). Each suction cup embeds commercial optoelectronic components in a multicolour soft elastomeric matrix (Fig. 1). Analytical and numerical modelling (Methods) optimized the silicone stalk geometry (the main optical sensing domain) for high sensitivity at low forces (~400 mV N⁻¹ in the 0–2-N range) as well as minimizing noise from external light. Experiments confirmed the ability to detect both magnitude and direction of applied forces, infer attachment and detachment, estimate object weight and localize contact positions along the arm.

The adopted solution features high reliability, compact read-out electronics with minimal wiring, low power consumption and small dimensions, allowing integration into small suction cups without bulky rigid parts. This advances existing implementations³⁹, which require large diameters^{40,41,42,43,44}, bulky rigid components^{45,46,47,48,49,50}, complex wiring^51,52 or lack directional sensitivity^53,54,55,56.

The distributed tactile sensing feeds a bioinspired control architecture that integrates local reflex loops and PNS processing to coordinate adaptive grasping behaviours. Underwater experiments demonstrated the robot’s ability to detect and localize contact with several cylindrical or relatively small objects, and perform autonomous, natural-like grasping^20,57,58. By fusing data from multiple suction cups, the arm can recognize the posture of a cylindrical object with respect to its central axis and execute twisting-based grasping or trigger local suction reflexes through cup-level circuitries, achieving context-dependent interaction.

Results

Suction cup characterization

The sensorized suction cups are key enablers of the intelligent and adaptive behaviours demonstrated by our soft robotic arm. We characterize their responses through indentation and pull-off experiments to assess reliability and accuracy, as well as confirm their ability to discriminate directional contact.

During indentation tests, the cups demonstrated high reproducibility with negligible variations in performance. The measured forces exhibit high accuracy, with a linear relationship (R² > 0.9; Supplementary Table 1) between the applied force and signal output (Fig. 2a,b). This trend holds true across all cup sizes and environments, with no evident correlation between sensitivity and cup dimensions (Fig. 2c). The average force–signal slope across all the tested combinations is 394.9 ± 21 mV N⁻¹, with noise below 5 mV. No statistically significant difference is observed between the dry and wet conditions, or between 1-N and 2-N loading conditions (Supplementary Fig. 1a–c and Supplementary Table 2), indicating a stable and repeatable sensitivity with ~0.1-N accuracy. Although there is no evidence of a strong drift over 300 cycles (Supplementary Table 1), some hysteresis is observed at higher loads (Fig. 2b and Supplementary Fig. 1d). However, the honestly significant difference reveals that this effect is significant only at 2 N in wet conditions (Fig. 2d and Supplementary Table 3). This suggests the onset of a viscoelastic-like response for the silicone material, probably due to enhanced passive adhesion in water (increasing suction cup deformation during uplifting). Endurance tests showed a stable signal with limited drift over time (Supplementary Fig. 2), strengthening the reliability of the proposed sensing strategy over time.

Fig. 2: Optoelectronic sensor characterization.

a, Force–signal relationship gained from the indentation tests performed for 300 cycles on sucker M in dry conditions. b, Force–signal relationship gained from the indentation tests performed for 300 cycles on sucker M in wet conditions. c, Slopes of the regression models fitting the force–signal values obtained from the indentation tests for all the sizes and in dry and wet conditions at 1 N and 2 N. d, Hysteresis is calculated as the loop area averaged over 300 cycles. The appendices on the histogram bars indicate similarities and differences resulting from the honestly significant difference analysis. In a–d, indentation is performed by uniformly compressing the suction cups (S, M and L) over a planar surface with forces of 1 N and 2 N. The response, obtained as the sum of the three signals provided by the phototransistors, exhibits excellent linearity in all cases (R² > 0.9) and yields an average sensitivity of 394.9 ± 21 mV N⁻¹. e, Zoomed-in view of a single cycle of the pull-off tests reveals the behaviour of the signal and force. The example from the L cup in a dry condition, with a preload of 1 N and a suction pressure of 0.2 bar, is shown in the figure. f, Relationship between force and signal for the cycle in e, highlighting the different phases (also shown in e): (i) approach and vacuum engagement, (ii) indentation until a specific preload is reached, (iii) pulling, (iv) suction cup anchored to the surface and stretching and (v) detachment. The signal remains linear even at negative forces (during pulling). g, Slope values of the regression lines for the force–signal relationship obtained from the pull-off tests were grouped by the applied suction pressure. Results suggest that neither the cup dimension nor the preload or applied pressure notably alters the force–signal behaviour during pulling (Supplementary Fig. 3 and Supplementary Fig. 1e,f). h, Maximum pulling forces (minimum pulling force plotted in absolute value) at different pressures for all cup sizes and test conditions. i, Circular indentation tests performed at different rim depths (i) considering the suction cup diameter (d) and the object. The schematics show the positions of each sensing location relative to the object at three indentation instants, when each sensor is at its maximum sensitivity. The polar plot shows the signal values (${||}{\bf v}_{1\it{j}}{||},{||}{\bf v}_{2\it{j}}{||},{||}{\bf v}_{3\it{j}}{||}$) from the three sensors (coloured differently) along the contact direction for three depths, corresponding to distinct colour grades for the sucker S. In the figure, O_S is the coordinate system of the suction cup and O_xy is the coordinate system of the disc. The contact direction is defined through a weighted vectorial sum based on the three signal values (Methods). j, Absolute average error in degrees between the actual and estimated contact directions obtained for the three dimensions at different indentation depths. For the box plots in d, g and j, central line and bottom/top edges indicate median and 25th/75th percentiles, respectively; whiskers extend to non-outlier extreme data; and mean values are overlaid through crosses. Sample size (n) and replicates (nr) are as follows: n = 3, nr = 300 in d; n = 3, nr = 10 in g; n = 1, nr = 180 in j.

In a single cycle of the pull-off experiments, several phases can be distinguished (Fig. 2e,f): first, the suction cup approaches the object; on contact, the vacuum generates suction, causing the force measured by the load cell to become negative, along with the signal perceived by the sensorized suction cup. The suction cup continues to compress until it reaches the set precompression value (1 N in Fig. 2e). At this point, the pulling phase begins, during which both force and suction cup signal decrease and then become negative. As the suction cup reaches the maximum pulling force, it stretches and is suddenly released. In particular, the pulling force exhibits an almost-linear behaviour, with a negative signal value during pulling. The linear behaviour remains consistent as the vacuum pressure increases (Fig. 2g). However, the maximum achievable force (absolute value of the negative force) increases with both cup size and suction pressure (Fig. 2h and Supplementary Figs. 3 and 1e,f). The negative signal values can be leveraged to identify an object being gripped and estimate its weight, following the linear relationship between signal and force (Supplementary Fig. 3). Results show an average sensitivity of 472.3 ± 47 mV N⁻¹ in pulling force, close to that in compression at 394.9 ± 21 mV N⁻¹.

Directional contact tests (Fig. 2i) showed that the difference between the computed contact position (defined through a weighted vectorial sum of the phototransistors read-outs; Methods) and the actual one was minimum (maximum error of <18° and absolute average error of <8° among the three different sized suction cups) when the contact happens in the proximal part of the rim (indentation equal to d/3, where d is the suction cup diameter), as the contact area is slowly increased (for example, maximum error of <40° and absolute average error of <22° when the indentation is equal to 2d/3) (Fig. 2j). As the contact area further increases, the ability to discriminate the direction of contact diminishes because the contact directional hints from the photoresistors become relatively marginal.

Soft arm architecture and control

The soft conical arm (Fig. 3a) encloses one straight tendon (T3) along the dorsal side, opposite to the suction cups, and two tendons (T1 and T2) ventrally routed in a crossed configuration (Methods). The dorsal tendon enables in-plane dorsal bending (T3 pulled; T1 and T2 released), whereas the ventral tendons enable in-plane ventral bending (T1 and T2 equally pulled; T3 released), as well as clockwise twisting (T2 pulled more than T1; T3 released) and counterclockwise twisting (T1 pulled more than T2; T3 released). With this tendon arrangement, the arm spans the configuration space illustrated in Extended Data Fig. 1, with respect to its base. The arm configuration space is then obtained through a sweep by 360° around the straight arm axis, through a rotational joint located at the arm’s base. External loads, such as forces exerted during grasping, cause configuration shifts (for example, relative to the load-free poses shown in Extended Data Fig. 1) due to the soft structure’s inherent compliance (Supplementary Fig. 4). Nevertheless, our arm is capable of grasping and bearing loads up to ~500 g (Supplementary Fig. 5).

Fig. 3: Octopus-inspired soft robotic arm architecture and control.

a, Prototyped arm with a schematic of its physical components: peristaltic water pumps, tendon pulleys and motors and the soft arm with a close-up view at its base showing the arm rotational joint about its (straight pose) axis, the three tendons, the suction channels and the electronic and communication wires entering the arm. Four large (L), four medium (M) and two small (S) artificial suckers were integrated into the arm and grouped into four suction channels. Internally, the soft arm comprises two tendons arranged on the ventral side (green (T2) and blue (T1)) and one tendon on the dorsal side (yellow (T3)). The sensing board of each suction cup is connected to a connection board. Together, they constitute, by analogy with the natural model, the local circuitry formed by the SG and the BG. Multiple connection boards are arranged in series and connected to a master board that embeds the peripheral control (PNS) and interfaces with an external PC for debugging. b, Octopus neural circuitries, as described in the literature⁵⁷. Elements of the LNC and the peripheral neural system are highlighted and mapped onto the physical architecture in the artificial arm through dashed rectangles. c, Proposed control scheme for the robotic arm implements multiple LNC control layers running in parallel, one for each suction cup, and a single, hierarchically superior PNS layer. The LNC layer detects contact, determines the contact direction and triggers local suction (each LNC operates independently and in parallel with the other LNCs and the PNS). The PNS layer collects contact direction information from all suction cups and defines the grasping strategy, overriding the LNC’s behaviour, if needed. A compact pseudocode of the main operations performed by the PNS layer is presented in the box, along with a schematic illustrating how different grasps are discerned. PID, proportional–integral–derivative.

The control of the soft arm is inspired by the hierarchical organization and distributed nature of the neural circuitry in the octopus, featuring an SG and a branchial ganglion (BG) on each sucker^7,20,57,58 (Fig. 3b). Similar to the biological model, the physical architecture of our arm features sensory-innervated suction cups, each associated with an SG (sensing board) connected to a BG (connection board), thereby forming an LNC (Fig. 3a,b). In our system, unlike the octopus, information from couple of suction cups converges into individual BGs distributed along the arm to minimize encumbrance (Methods). The current implementation does not include CNS functionality or inter-arm communication. We rather focus on a single arm and imitate the PNS by coordinating the distributed sensory data from all suction cups via the master board.

The embedded control has a hierarchical structure. The top PNS layer connects to multiple lower LNC layers, one per suction cup (Fig. 3c). The parallel LNCs implement a reflex behaviour that triggers local suction when at least one phototransistor signal, normalized on s_max, exceeds a sensitivity threshold (${v}_{{a}_{{\rm{th}}}} > 0$). Simultaneously, sensory information is sent to the PNS layer that determines the grasping strategy (which, in turn, can remodulate suction activation/deactivation with authority, as detailed below) and triggers tendon actuation.

We defined a set of simplified, non-optimal grasping strategies for proof-of-concept demonstrations, based on the object orientation angle β relative to the arm longitudinal axis (in straight pose), estimated through the active suction cups. More specifically (Methods provides additional details): (1) if at least one of the last three peripheral suction cups (SCs; SC 8–10) detects contact, then dorsal bending is triggered; (2) if only a single cup (from SC 1 to SC 7) establishes contact, ventral bending is triggered; (3) if more than one suction cup detects contact, either ventral bending or clockwise/counterclockwise twisting is performed depending on the estimated pose β of the object; (4) suction yet no tendon actuation is commanded when the object is almost aligned with the longitudinal axis of the arm. In parallel with local reflexes, the PNS processes information (namely, the local contact direction) arriving from the suction cups and establishes a grasping strategy after an observation time window (Δt). The definition of the grasp may also override local circuit behaviours. Specifically, when performing dorsal bending, the PNS keeps active only the group associated with the suction cup that detected the contact within Δt. In all other cases, the PNS takes control over the LNCs and activates suction in all cups (Fig. 3c). This is because we have observed this strategy to be more effective, especially in underwater conditions in which the objects might float or slip away.

Detection of contact, force and direction by the suction cups in the arm

The fully integrated soft arm was suspended in the air to test readings from individual suction cups (Supplementary Video 1). Encapsulation of the suction cups within the arm did not affect signal reliability (Fig. 4a). The system clearly distinguishes which suction cup is being touched, as each signal deviates from its baseline. Contact force can be recognized from the cumulative signal of the phototransistors within the cup, and different contact directions can also be identified.

Fig. 4: Arm sensory feedback.

a, Results of the sensing capability with the suction cups integrated in the arm. The bottom graph displays the cumulative signal (sum of the three phototransistor outputs) from each suction cup. The top inset graphs show the individual signals generated through the stimulation, as in the frame above. Different cups (in this test: SC 1, SC 3, SC 5 and SC 8) have been stimulated, resulting in increased output signals. SC 4 has been stimulated with a finger, increasing pressure over three distinct touches. The cumulative signals demonstrate an increase in signal output, which can be converted to the force exerted by applying the linear relation established during characterization (Fig. 2). Forces at peak values: 0.23, 0.45 and 0.73 N. Finally, a single suction cup (SC 2) has been contacted in three different locations around the rim (as depicted in the top-right diagrams), and the contact directionality can be discriminated by following equations (1) and (2). b, Two different suction cups have been stimulated at opposite sites to demonstrate the ability of the arm to discriminate contact direction and reorient towards it. Numbers in the circles refer to the first instant of contact. The sequence of arm rotation is shown for the first two contacts. c, The arm has been contacted at its distal suction cup with an object: a plastic cup containing a 100-g calibrated weight. Graphs show the signals from the phototransistors acquired by the contacted suction cup, cumulative signal, estimated weight of the object and tendon actuation.

We used this information to control the arm orientation during underwater tests, adapting the above algorithm accordingly (Supplementary Video 1). If only one of the proximal cups (SC 1–6) is contacted at the sides (proximal to 0 or ${\rm{\pi }}$ relative to the cup reference frame), then the arm reorients following that direction until the contact ends or additional cups contact. Two suction cups were individually contacted twice at opposite sites. Tests show the correct detection of contact directions and continuous arm rotation as contact is maintained (Fig. 4b).

We also tested the feasibility of estimating the weight of an object suspended through suction (Fig. 4c). A 55-mm-diameter, 64-mm-high box filled with water was used as a container for a weight of 100 g, giving an apparent underwater weight of about 85 gf. When the last suction cup detected contact (Fig. 4c, point 1), the local circuit correctly activated suction for the fourth group. Once adhesion was established (Fig. 4c, point 2) and before actuation began, the signal stabilized, allowing estimation (about 72.5 gf) of the apparent weight using the previously characterized linear relation between cumulative signal and pulling force (slope, 430 mV N⁻¹; Supplementary Fig. 1f).

Interestingly, during suction, the signals exhibit a fast, transient drop that resets the observation window used by the PNS to determine the pose of the object (Supplementary Video 2). However, local reflexes promptly activated local suction at first contact, maintaining adhesion with the object throughout the dynamics, ensuring a successful grasp.

Identification of object posture and underwater grasping

The effectiveness of the hierarchical, behaviour-based control to drive natural-like movements in the soft arm was tested in underwater conditions by placing an object in contact with the suction cups at different inclinations (Supplementary Video 3). Sensor readings allow a clear identification of the object’s contact with individual suction cups (Fig. 5a (black, solid line) and close-up views (dashed black rectangle) for cups SC 3 and SC 4). When an artificial sucker detects the contact $(\exists i\,{\rm{s}}.{\rm{t}}.\,{||}\bf{v}_{{i}{j}}{||}/{s}_{\max } > {v}_{{a}_{\mathrm{th}}})$, it immediately activates suction for its group (Supplementary Video 4), facilitating adhesion to the object and subsequent grasping.

Fig. 5: Arm sensing and control loop in underwater grasping.

a, The arm has been contacted with a cylindrical object to test the control strategy. Graphs display the signal output from each suction cup over time, along with the corresponding tendon actuation (above). When a suction cup detects contact (one of the signals from phototransistors ||$\bf{v}_{{i}{j}}$|| normalized on s_max rises above ${v}_{{a}_{\mathrm{th}}}=0.25$), it triggers suction for its suction group, and the peripheral control evaluates and stores, for a time lag of Δt ≈ 4 s, the object inclination β following equations (3) and (4). At the end of this period, the mean β is used to trigger tendon actuation. Close-up views of the signals in the period of tendon actuation show a characteristic linearly descending trend in artificial suckers not in contact with the object. b, The arm was initiated with a slow ventral movement (point 1). Although bending, a cylindrical object was placed in front of the arm, and the contact was correctly detected (point 2). The arm continued to bend until the time lag of Δt (~4 s) had elapsed, after which, the actuation was automatically corrected (point 3) to perform a twisting and grasp the object (point 4). Close-up views highlight baseline deviations in the suction cups that are in contact during the observation window (dashed black rectangle), and in both contacting and non-contacting suction cups during the subsequent actuation window (dashed red rectangle).

It can be noted that during actuation, all signals slightly deviate from their baseline, even in those suction cups not contacting the object (close-up views in Fig. 5a (dashed red rectangle)). This phenomenon is due to actuation-sensing cross-talk that, however, only occurs after the PNS has defined the grasp, which remains unaffected. In particular, signals also exhibit a characteristic decreasing trend in suction cups that do not establish contact with the rim during actuation (Fig. 5a, cups 1 and 4–10), and the signal intensity increases with tendon tension. Differently, when the arm contacts an object, the suction cups at the contact site systematically show a positive variation, up to reaching positive values (Supplementary Figs. 6 and 7).

To further validate the reliability of our system, we challenged the embedded control to detect and grasp an object as the arm was already in motion, for instance, during (ventral or dorsal) bending (Supplementary Video 4). A continuous, slow arm movement until contact is detected and can be exploited as a basic search strategy, potentially useful for exploring delicate/unknown environments. Figure 5b shows the results of the test in which the arm made contact with a cylindrical object and performed a ventral bending. The graphs show the instant of contact detection (~8 s) and the change in actuation once the PNS defines the grasp (~12 s). In particular, signals during slow ventral bending maintain their characteristic baseline pattern, without an evident effect of cross-talk between sensing and actuation. Once the grasp is triggered, signals display the characteristic linearly decreasing behaviour in non-contacting suction cups. Importantly, it is still possible to discriminate which suction cup comes into contact with the object during actuation by observing the rising of the signals that deviate from the linear trend and become positive (for example, in dashed red rectangles in Fig. 5a (SC 2) and Fig. 5b (SC 4)).

The arm with its embedded control strategy demonstrates effectiveness by successfully detecting and autonomously grasping multiple (similarly) cylindrical or relatively small objects at various contact positions, in relatively complex, underwater and unknown scenarios with varying combinations of movements and search tasks (Fig. 6 and Supplementary Video 5 show a larger selection).

Fig. 6: Examples of underwater grasping of different objects.

a, The arm accomplishes a search task using the rotary joint at the base to explore the environment horizontally. When contact with the external object (a glass bottle) is detected, the search task stops, the PNS computes the object pose and the grasp is triggered. b, The arm, in a vertical orientation, contacts a cylindrical object and properly triggers ventral bending. c, The distal suction cup is contacted with an empty cup, and only local suction is triggered with dorsal bending. d, The arm, positioned horizontally with its ventral side facing downwards, explores the environment through ventral bending. When detecting a contact (an artificial seastar is used in the example), local suction is activated, and the PNS opportunely computes dorsal bending to retrieve the object.

Discussion

This work presents an octopus-inspired soft robotic arm equipped with sensorized suction cups that provide exteroceptive tactile feedback and support peripheral decentralized control architectures. By embedding optoelectronic sensors directly into the stalks of soft suction cups, we developed a distributed sensing platform that endows the robotic arm with autonomy and adaptability during grasping tasks. The suction cup design and material were optimized to achieve high sensitivity at low contact forces, aligning with the mechanical properties of the soft arm, which inherently limit the maximum contact force, especially at its periphery. This design enables the reliable detection of gentle interactions at low contact forces through the integrated suction cups, ensuring overall softness and compliance. This makes it suitable for delicate manipulation and potential underwater applications.

The sensors are located in the suction cup stalk, which has a constant dimension across different cup sizes. This ensures that the signal is independent of the suction cup diameter. The characterization of the suction cups shows high linearity between sensor outputs and applied force, allowing the system to detect both contact and its directionality with high accuracy (<18°). Higher accuracy could be reached by increasing the number of LED–phototransistor pairs in stand-alone sensing applications. However, in this work, we used three pairs, as this is the minimum for directional discrimination and maintaining scalability, integrability, energy efficiency, data acquisition speed and a compact mechanical integration within the soft arm. Preload effects were also negligible, further simplifying sensor calibration and interpretation. Hysteresis effects, which tend to increase under wet conditions and at higher forces (around 2 N; Supplementary Fig. 1d), are not a major concern within the operational force range, as long as the fidelity of contact detection is maintained. The characterization demonstrates the comparability of results across different suction cup implementations, thereby making individual cup calibration unnecessary and standardizing signal interpretation. The proposed suction cup advances the state of the art (Supplementary Table 4), which assesses its scalability and integrability into soft, slender bodies. Most existing sensorized suction cup solutions rely on external processing units (for example, optical demodulators, pressure transducers or camera-based image processing), which limit compact integration within slender, soft bodies. Several approaches also depend on machine learning-based data interpretation with centralized processing. By contrast, the proposed optoelectronic architecture integrates both sensing elements and read-out electronics directly within the arm, enabling miniaturized suction cups (diameter, 12–20 mm) compatible with soft continuum manipulators. Force is estimated through a direct linear mV–N calibration, whereas the contact direction is reconstructed in real time via a vector resultant of the three optical channels, avoiding complex external processing. The use of commercial optoelectronic components in the proposed design also contributes to environmental robustness and adaptability, in both dry and underwater conditions, allowing to contain developmental times and costs as well.

With regard to the arm, the existing cross-talk between tendon-driven actuation and sensory output (Fig. 4) does not represent a detrimental limitation. First, it is inherent to soft-bodied systems, including biological models, in which a complete decoupling between external and internal forces acting on embedded sensors never exists. Moreover, the observed signal deviations from the baseline are highly repeatable across cycles. In particular, they remain consistently negative and increase in magnitude with tendon tension, whereas when the arm contacts an object, the suction cups at the point of contact exhibit a clear positive response (Supplementary Figs. 6 and 7). This substantial difference can be used to infer proprioceptive feedback in future implementations of the arm. Moreover, in perspective, transitioning to neuromorphic computing for the joint interpretation of exteroception and proprioception could enable the system to emulate natural neural circuits more closely.

To ensure the long-term reliability of our arm, all components were selected and designed for mechanical durability and stability in wet environments. Specifically, we also considered potential failure modes to choose and test individual components and subsystems (Supplementary Table 5) and have reached a soft arm that displays high endurance (Supplementary Fig. 8). Building on these initial design choices, a broader and more systematic analysis of potential failure modes (for example, through a failure modes and effects analysis approach) could be carried out for future implementations, based on an extended characterization of the arm’s tasks (for example, as part of a larger system) and its actual working environment.

Beyond sensing, the system architecture supports modularity and scalability. Owing to the scalability of the suction cups, their number and configuration along the arm can be easily varied. This feature not only provides flexibility in distributing attachment points but also in tuning the distribution of exteroceptive locations. This is an important consideration that enriches the design domain for both hardware and control of the soft robotic manipulator, enabling it to span various application needs. This work proposes a compact design solution for soft robotic arms that integrates sensing and computation, enabling a tight coupling between sensing and actuation at the local level and supporting arm-level decision-making for activating diverse grasping strategies. In particular, the integration of local reflex behaviours (executed independently and in parallel) with higher-level decision-making enables a prompt response to contact. Local suction activation stabilizes the initial contact, which is then followed by intentional tendon actuation triggered by the peripheral circuitry, shaping and further stabilizing the grasp. By processing tactile signals locally at each suction cup and transmitting only contact direction information to the PNS, the distributed architecture limits communication bandwidth requirements and reduces the central computational load. This enables the system to scale to a larger number of suction cups (and potentially multiple arms) without compromising the acquisition rate or response speed, unlike centralized approaches that require the continuous transmission of raw sensory data and have higher computational load on the central controller.

Our robotic arm demonstrates the feasibility of implementing distributed, hierarchically structured control of octopuses within a compact design suitable for both air and underwater environments. All underwater experiments were performed in a fully submerged environment, as the arm is designed to operate underwater, where its near-neutral buoyancy (density, ~1.07 g cm⁻³) enables stable and multidirectional movements. We are aware of the simplification of assuming similar cylindrical or small and lightweight objects in non-floating conditions for our grasping experiments, which nonetheless allowed us to demonstrate the key characteristics of the system. A wider set of grasped objects will be addressed in future studies, where even the load-bearing capacity of the arm could be optimized, based on application-specific targets and constraints. Furthermore, tendon sensing feedback, which was not needed for the purposes of the present study, could be incorporated in future application-specific embodiments to develop more refined control strategies.

Despite the simplifications in the current implementation, the proposed distributed, embedded sensing and control approach enables the soft, continuum robotic arm to autonomously respond to contact events without relying on external, centralized control. This represents an important step towards more intelligent, adaptable and resilient soft robots.

Methods

Suction cup design

The proposed sensorized suction cup exploits optical intensity measurements to detect tactile stimuli. A miniaturized electronic board (diameter, 7.3 mm) composed of three LEDs interspaced by three phototransistors and circularly assembled at 120° is integrated into the stalk of the suction cup and directed towards a reflective surface (Extended Data Fig. 2). Suction cups have been realized with a small (S = 12 mm), medium (M = 16 mm) and large (L = 20 mm) diameter (Fig. 1), whereas preserving a constant stalk diameter (10 mm) for the three cup sizes.

The suction cups are made of Ecoflex 00-50 silicone rubber. Its compliance (in the 150–400-kPa range⁵⁹) ensures sufficient deformability, allowing conformation to rough surfaces and achieving good sensitivity under low contact forces (<1 N) (Supplementary Fig. 9a). This high compliance also enhances encapsulation and gripping force. Moreover, the selected elastomer is relatively transparent in the visible and near-infrared ranges (transmittance, 50%–20% for thicknesses of 1–5 mm; Supplementary Fig. 9b), allowing light to travel through the suction cup body with low attenuation, thereby reducing both light intensity and power consumption. A reflective surface, positioned at a specific distance from the board, directs the LED light towards the phototransistors. This surface is obtained by adding a white colourant to the base silicone (Supplementary Fig. 9c,d). This simplifies fabrication and preserves the cup’s softness and mechanical integrity. A black silicone layer (obtained by incorporating black colourant into the base material) was added to shield the internal waveguide from external light, thereby minimizing interference and saturation from ambient illumination (Supplementary Fig. 9c). Additionally, a differential measurement strategy was adopted to mitigate environmental light effects further: the signal acquired without LED illumination is subtracted from the signal with LED illumination, effectively removing the background contribution (Supplementary Fig. 10).

Optical sensor modelling

Starting from the optoelectronic board dimensions, we initially chose a 10-mm outer diameter for the stalk and a 2-mm diameter for the central vacuum channel. Moreover, to achieve the sought-after miniaturized implementation, we assumed relatively thin black silicone shielding: 0.1 mm for the central channel and 1 mm for the outer boundaries. Furthermore, we set the white silicone thickness to 1 mm for axial compactness and ease of manufacturing. On the basis of the above starting choices, the considered optical working thickness (denoted by h in Extended Data Fig. 2), which is the axial distance between the reflective surface and the external surface of LEDs and phototransistors, remained the free design parameter to be assigned by maximizing the sensitivity of the optical read-out.

To this purpose, we used analytical and numerical modelling (Supplementary Text). A simplified two-dimensional analytical model (Extended Data Fig. 2d) was first developed for a single LED–phototransistor pair to study competing effects at smal h l values, where short, highly inclined rays reach the phototransistor. Shorter paths reduce intensity loss through the translucent medium, whereas higher inclinations attenuate light due to the LED and phototransistor angular response (Supplementary Fig. 11a). The model predicted these competing effects around h ≈ 0.5 mm (Supplementary Fig. 11b), with both deposited power and sensitivity decreasing monotonically for larger h. Representative trends are shown in Extended Data Fig. 2d.

To leverage the cues provided by the two-dimensional approximation, we introduced a three-dimensional (3D) numerical model (Extended Data Fig. 2e) that incorporates the actual geometry, material properties and LED/phototransistor characteristics. The 3D model confirmed the decreasing trend of deposited power and detection sensitivity for h > 0.7 mm. To ensure monotonic signal response under deformation, an optical thickness of 1 mm was selected, above the non-monotonic transition (~0.5 mm), maintaining high sensitivity, as a 2-N load yields ~0.3 mm shortening within the monotonic range.

Computation of the contact direction

On contact, signal intensity for each phototransistor of a sensorized suction cup was the maximum when the contact was oriented towards the sensor, and decreased as it became misaligned, reaching a minimum when oriented in the opposite direction. Owing to this characteristic, we defined a vector v_ij oriented towards each phototransistor i in suction cup j, having a magnitude ${||}\bf{v}_{ij}{||}$ equal to the signal intensity in that direction. Once normalized v_ij through the averaged maximum signal towards each phototransistor (s_max = 66.7 mV with 0.5 N of force) gained from cup characterization, we then computed the contact direction (s_j) and the contact angle (α_j), as obtained by its four-quadrant inverse tangent:

$${\bf{s}}_{j}=\mathop{\sum}\limits_{i=1}^{3}{\bf{v}}_{ij},$$

(1)

$${\alpha }_{j}={\tan }^{-1}\frac{{s}_{j{\rm{\_}}\,y}}{{s}_{j{\rm{\_}}x}}\,.$$

(2)

Estimation of object pose and selection of grasp

Object orientation relative to the arm longitudinal axis in the relaxed configuration is approximated by summing the contact direction vectors (s_j), from each suction cup, after reflecting those in the lower half-plane to align with the upper quadrants, yielding the cumulative contact direction (r):

$${\bf{r}}=\mathop{\sum }\limits_{j=1}^{10}{\bf{s}}_{j}.$$

(3)

The object inclination with respect to the central axis of the arm is then obtained as the angle of the normal to $\vec{r}$:

$$\beta =\left({\tan }^{-1}\frac{{r}_{y}}{{r}_{x}}+\frac{{\rm{\pi }}}{2}\right)\,\mathrm{mod}\,{\rm{\pi }},$$

(4)

which gives β = 0 if the object is oriented perpendicularly to the arm axis and β = 90° if they are aligned. In the control, a time lag (Δt) is imposed to allow the system to reach a relatively stable contact, during which signals also stabilize (Fig. 5a). Such an observation time window can be shortened or lengthened depending on specific application scenarios. The object inclination is then evaluated by taking the average of the β angles obtained using equation (4) along Δt. Information gathered from each suction cup, their individual activation (signal rising from an unstimulated condition) and the aggregated information concerning the object inclination define the grasping strategy and trigger tendon actuation. The grasp is defined by combining information about the cups in contact and β. If at least one of the last three peripheral suction cups (SC 8–10) detects contact, dorsal bending is triggered. If only one of the other suction cups establishes contact, ventral bending is triggered. If more than one suction cup detects contact, a threshold strategy is followed. It has been imposed to bend ventrally if $\beta > 160^\circ$ or $\beta < 20^\circ$. A clockwise twisting is performed if $20^\circ \le \beta \le 80^\circ$, and a counterclockwise twisting if $100^\circ \le \beta \le 160^\circ$. If more than four cups perceive contact or $80^\circ < \beta < 100^\circ$, the object is probably aligned with the arm central axis; thus, no tendons are actuated, but only suction is performed.

Electronics design, fabrication and low-level control

The sensor board, located in the stalk of each cup, comprises three LEDs (15406094BA400A, 940 nm, Würth Elektronik) connected in series to provide them with the same current and ensure uniform illumination in the suction cup. The phototransistors (1540601NBA500, 940 nm, Würth Elektronik) were connected in a common collector configuration. LEDs in the suction cups are powered by a 5-V digital pin with a 1.8-kΩ resistor, resulting in approximately 1 mA of current. A 4.7-kΩ resistor is connected to the emitter of the phototransistor that amplifies the photocurrent (Supplementary Fig. 12). The resulting sensor board was designed with a circular shape (diameter, 7.3 mm) and an internal hole (diameter, 2.1 mm) for vacuum actuation. Six pads on the back side were used to wire the board to the connection boards. Both sensor and connection boards were fabricated in FR-4 with a thickness of 0.5 mm (Eurocircuits NV).

Five connection boards (11.7 × 10.3 mm²; Supplementary Fig. 12) collect data from pairs of suction cups. Each connection board comprises a microcontroller (CY8C4045LQI-S411, Infineon Technologies) that serves as a communication node and data acquisition device. Drawing inspiration from the spiral arrangement of the octopus axial nerve cord, which permits elongation without tissue damage¹⁸, we used coiled wires for a four-line bus to connect the electronic boards along the arm to a master board. This approach minimizes mechanical interference with arm movements and prevents wire breakage during local extensions (Supplementary Fig. 13). The bus brings power (5 V and GND) and an I2C (SDA, SCK) interface.

To achieve a high data acquisition rate of 200 Hz, the microcontroller operates at 32 MHz. Data transfer occurs via I2C communication at 400 kHz, initiated by the master board located at the base of the arm (Fig. 3a). The master board sequentially reads data from the five boards, each assigned a unique I2C address. The data sampling rate is synchronized with the data transfer process; sampling begins immediately after each data transfer completes, ensuring synchronization; therefore, data are always ready when requested by the main controller.

The master board is composed of three microcontrollers (PIC32MX150F128B, from Microchip): one manages communication with both external PC and connection boards, whereas the other two control the four two-channel motor drivers (LV8548MC, from ON Semiconductor) used for the three tendons, the rotating mechanism and the four peristaltic water pumps (WPM447, Velleman). The operating parameters of the octopus arm are selected through a custom graphical user interface developed in VB.NET, which connects to the master board via USB. The same interface also receives feedback from the arm, allowing the real-time monitoring of sensory data.

The connection to the external PC is used solely for debugging, to visualize sensory information, and to control atomic actions when the execution mode is ‘idle’ (during testing and debugging). When the arm is set to ‘autonomous’ via the interface, it operates entirely through the embedded control running on the master and connection boards, transmitting only sensory data and internal status to the PC for user visual feedback. The code for the embedded control is available in the online repository⁶⁰. All experiments demonstrating grasping abilities, including search, contact, posture identification, grasping definition and actuation, are performed via the embedded control.

Fabrication of suction cups

Ten sensorized suction cups were integrated into an octopus-inspired soft arm (Fig. 3). The overall fabrication process can be considered a two-step process, where the first step involves fabricating the suction cups (Extended Data Fig. 3a), as outlined in this section, and the second step involves realizing the arm (Extended Data Fig. 3b), as described in the following section.

To realize the suction cups, silicone was mixed for 60 s using a THINKY ARE-250 mixing and degassing machine (THINKY USA), followed by a further degassing step of 3–5 min using the S-26P-NL vacuum degassing system (Easy Composites). The sensor board is fabricated in FR-4 with two sacrificial supports (easy to break and remove) to permit its placement inside the moulds during suction cup fabrication. Six wires (enamelled copper wire, 0.15 mm) were soldered to each board to power the LEDs and manage acquisition from the phototransistors. The boards are secured in a two-part polytetrafluoroethylene mould (fabricated by CNC machining) coupled by five screws (Extended Data Fig. 3a (top), steps 1–3). A third part is used to manufacture the white reflective layer (Extended Data Fig. 3a (bottom), steps 1–3). The mould is planarly placed and filled with silicone (EcoFlex 00-50, Smooth-On) mixed with 1% white colourant (Silc Pig white, Smooth-On). After degassing, the surface is leveled with a smooth scraper (Extended Data Fig. 3a (bottom), step 2) and left to cure at ambient conditions for 45–60 min (Extended Data Fig. 3a (bottom), step 3). This period has been verified as sufficient to prevent mixing with the subsequent silicone layer and ensure good bonding to prevent delamination. After that, the two moulds are fixed together (Extended Data Fig. 3a, step 4), and the non-coloured silicone is mixed, poured into a syringe and degassed before being injected into the mould (Extended Data Fig. 3a, step 5). The injection allows the silicone to infiltrate the bottom part and release it from the holes located at the top of the mould, ensuring the removal of air bubbles that can otherwise significantly impact the optical performance of the sensor. After curing, the fabricated cylinder is gently extracted from the mould (Extended Data Fig. 3a, step 6), preserving the sacrificial supports, and then inserted into a new three-part mould to give the suction cup shape. A black colourant (Silc Pig black, Smooth-On) is mixed with the same silicone, and, as in the previous step, it is poured into a syringe and degassed before being injected into the mould (Extended Data Fig. 3a, step 7). After curing, the final suction cup is extracted and ready for integration into the arm (Extended Data Fig. 3a, steps 8 and 9).

Fabrication of octopus-inspired soft arm

The arm geometry consists of a cone with a base radius of 20 mm and a tip radius of 4.1 mm, with a length of 410 mm. A mould was 3D printed using ProX SLS 6100 (3D Systems) with DuraForm ProX PA, a commercial polyamide grade, featuring ten cylindrical slots designed to hold the suckers during the initial positioning stage. Before casting, the ten suckers are connected in pairs to five control boards by means of coiled wires, and the control boards are connected in a daisy chain configuration (Extended Data Fig. 3, steps 10 and 11). In parallel, silicone tubes with inner and outer diameters equal to 0.76 mm and 1.65 mm, respectively, are connected by L- and T-barbed 1/16′ connectors to create four independent suction circuits, each linking four, three, two and one sucker, respectively (Extended Data Fig. 3b, step 11).

These four channels were integrated into the arm to drive the suction cups via external peristaltic water pumps. Although the independent actuation of each cup is ideal, such an approach would require too many tubes along the arm, reducing the available space and increasing the stiffness. To balance functionality and flexibility, the cups were organized into four groups, each driven by a single pump.

Moreover, three commercial fishing wires (Spiderwire DuraBraid) with a diameter of 0.35 mm were arranged in the mould to act as tendons and control arm motion. Two of them were arranged in a piecewise-linear configuration, symmetrically with respect to the longitudinal plane of the arm, and the third one was simply linear, decentred towards the dorsal side (Extended Data Fig. 3b, step 11). Silicone tubes with inner and outer diameters of 0.76 mm and 1.65 mm, respectively, were used as tendon guides to prevent tearing of the cured silicone during regular arm use due to tendon sliding. The tubes were passed through eyelets on three 3D printed supports, intended to be positioned at the cone base and at distances of 250 mm and 343 mm from it. A double knot was made at the tendon anchoring points to hold the wires during their pulling. Each 3D printed support was provided with two thin sacrificial protrusions, designed to be positioned in the designated mould slots (Extended Data Fig. 3b, step 11).

During assembly, the suction cups and their boards, suction channels and tendons were carefully positioned into the mould (Extended Data Fig. 3b, step 12). The tendons were put under slight tension, and special care was taken to ensure that all components remained embedded within the final conical shape.

The final fabrication steps included mould closure (Extended Data Fig. 3b, step 13), the pouring of silicone (EcoFlex 00-30, Smooth-On) mixed with 1% red colourant (Silc Pig red, Smooth-On), and the extraction of the final arm after 4 h of curing at room temperature (Extended Data Fig. 3b, step 14). Views of the fabricated arms are shown in Supplementary Fig. 14.

To test its reliability and endurance, we performed an extended actuation test of 10,000 cycles under a 14-N tensile load, sufficient to fully contract the tendon (10 cm) in air.

Actuation

The tendons are actuated by d.c. motors (Faulhaber 1717T012SR) equipped with embedded encoders and gearboxes with an 81:1 reduction ratio, operating in the position-control mode. Each motor drives a tendon via a pulley (Ø, 3 mm), enabling accurate, repeatable tendon displacement. In addition, we used a rotational joint (Fig. 3a) at the base of the arm, consisting of a 3D printed circular base plate onto which the motors are mounted, with circumferentially distributed teeth along its outer rim. Rotary motion is provided by an external motor (Pololu 37D, 50:1) fitted with a smaller pinion, allowing rotation at up to 45° s⁻¹. Four roller-guide assemblies are placed around the rotating base to ensure vertical and radial constraints.

Suction cup characterization

We conducted indentation and pull-off experiments using a 3D printed polylactic acid planar surface anchored to the slider of a custom setup mounted on an optical table, to evaluate the sensorized suction cup mechanoreception capabilities in air and underwater conditions. Performances were assessed by measuring the contact force, gripping force and contact direction under various configurations with the three cup diameters.

The suction cups were subjected to 300 indentation cycles at 1 N and 2 N to verify accuracy, hysteresis and signal drift, as well as to assess the repeatability and endurance of the sensor within the proposed embodiment (Supplementary Video 6). The outputs of the three phototransistors were summed to produce a cumulative signal for analysis.

To demonstrate the durability and reliability of the suction cup, we conducted 20,000 indentation cycles on the M-type suction cup (Supplementary Fig. 2).

Ten cycles of pull-off experiments were performed by approaching the 3D printed polylactic acid surface and activating different vacuum pressures (0.1, 0.2 and 0.4 bar) (Supplementary Video 7).

For the indentation and pull-off experiments, the suction cups were first attached to 3D printed rigid supports using a silicone adhesive (Sil-Poxy, Smooth-On), ensuring that the sensing area remained exposed. Each prepared suction cup was connected to a six-axis load cell (Nano17, ATI) via a 3D printed adaptor. The suction cup and load cell were moved by a vertical motorized linear stage (M-111 Compact Micro-Translation Stage, PI), and data were acquired through a data acquisition board (NI PCI-6225, National Instruments). A custom graphical user interface developed in VB.NET was used to record the data and control the linear stages.

To check the suction cup ability to detect contact direction, a rigid 10-cm-diameter disc was 3D printed (Prusa MK4S, Prusa Research) and indented using the same setup, with the disc positioned by two additional X–Y linear stages (model M-414.3PD, PI) to realize a circular path. The circular target object was contacted by the sensorized suction cup from all directions with an angular step of 15° and a compression force of 0.5 N (Supplementary Video 8). We repeated the experiments by contacting the suction cups at different positions, progressively increasing the contact area with the object (Fig. 3i). The intensities of the three phototransistor signals were evaluated independently.

Experimental validation

All autonomous grasping experiments were performed in fully submerged freshwater laboratory tanks (60 × 60 × 60 cm³ and 90 × 50 × 50 cm³) under static water conditions (no active flow), at 21 ± 2 °C and under ambient laboratory lighting supplemented with diffuse LED lights for video recording. No external perception systems, vision-based guidance or teleoperation were used during the tests.

Target objects included cylindrical and quasi-cylindrical bodies with varying diameters (in the range of 3–6 cm) and inclinations, as well as smaller (characteristic size of a few centimetres) and irregularly shaped items. Object weights ranged from approximately 0.1 N to 5 N. Objects were positioned, with different orientations, along the arm’s longitudinal axis to estimate contact location and object relative pose, to trigger different autonomous grasping. Grasp selection and execution were carried out exclusively based on distributed tactile sensing, according to the processing strategy implemented in the embedded hierarchical control architecture.

An overview of the experimental validation framework adopted in this study is summarized in Supplementary Table 6, where the performed tests are grouped into three categories aimed at assessing (1) the embedded sensing and signal processing capabilities, (2) the operation of the hierarchical control architecture under pre-planned conditions and (3) the robustness of the control in more unstructured and non-planned interaction scenarios.