PinchBot: Versatile Pinch-Based Robotics

• PinchBot is a robotics paradigm defined by pinch-based interaction, enabling precise, adaptive control in varied applications such as pottery shaping and tactile grasping.
• In pottery creation, a guided diffusion policy directs sequential pinch actions that gradually morph clay into a target 3D shape with measurable precision.
• Pinch-centric designs integrate specialized end-effectors, tactile sensors, and wearable assistive systems to enhance grasp stability and adaptivity under uncertainty.

PinchBot is a designation used in recent robotics literature for several systems organized around pinch-based interaction. In the provided corpus, the term most directly denotes a goal-conditioned robotic platform for long-horizon deformable manipulation in pottery creation, where a robot gradually sculpts clay through a sequence of pinch actions predicted by a guided diffusion policy (Bartsch et al., 23 Jul 2025). The same designation also appears in relation to pinch-centric end-effectors for linear parallel pinching and self-adaptive grasping, tactilely stabilized pinch-grasp controllers, a soft wearable exoskeleton for thumb-index assistance, and a planar manipulator built from pinched bistable tapes (Ding et al., 18 Oct 2025, Psomopoulou et al., 2021, Grønvall et al., 3 Jun 2026, Osele et al., 2021). Across these uses, the unifying theme is the treatment of pinching not merely as a grasp primitive but as a control, sensing, and mechanical-design principle.

1. Terminological scope and research landscape

In the provided literature, “PinchBot” does not identify a single canonical hardware stack. Rather, it spans several distinct research programs that share a pinch-centered manipulation paradigm.

Usage in the literature	Core function	Representative papers
Goal-conditioned pottery robot	Long-horizon deformable manipulation by pinch actions	(Bartsch et al., 23 Jul 2025)
Parallel gripper / adaptive hand	Linear pinching and self-adaptive grasping	(Ding et al., 18 Oct 2025, Guo et al., 15 Oct 2025)
Tactile pinch systems	Stable 3D pinching and few-shot imitation learning	(Psomopoulou et al., 2021, Mao et al., 2023)
Assistive or structural embodiments	EMG-driven pinch assistance; planar manipulator using pinched bistable tapes	(Grønvall et al., 3 Jun 2026, Osele et al., 2021)

A common source of ambiguity is therefore nominal rather than technical. The pottery system is a policy-learning platform; the SP-Diff and Hoecken-D Hand are end-effector mechanisms; the tactile systems are closed-loop grasping frameworks; SoftPINCH is a wearable assistive robot; and the pinched-bistable-tape device is a planar manipulator. The recurring scientific question is how pinch-based interaction can be made precise, adaptive, and robust under uncertainty in object geometry, compliance, or user intent.

This breadth also clarifies a frequent misconception: in the cited work, pinching is not restricted to rigid precision grasping. It includes deformable shape formation, self-adaptive enveloping, contact-aware re-grasping, EMG-triggered assistance, and joint formation by localized structural pinching.

2. Goal-conditioned deformable manipulation for pottery

The most explicit use of the name is the system introduced in “PinchBot: Long-Horizon Deformable Manipulation with Guided Diffusion Policy,” which addresses pottery creation as a highly multi-modal and long-horizon deformable manipulation task (Bartsch et al., 23 Jul 2025). The task is formulated around repeated pinch actions that slowly morph a block of clay toward a target 3D bowl shape.

The state at time $t$ is a 3D point cloud of the clay, down-sampled to $N=2048$ points. Each point cloud is obtained by fusing four fixed Intel RealSense D415 cameras and one end-effector-mounted D415 through color thresholding, extrinsic calibration, end-effector pose transforms, ICP alignment, and uniform sampling. Goal conditioning is defined by a target point cloud $C_{\text{goal}}\in\mathbb R^{N\times 3}$ of the finished bowl, encoded into a latent vector $g=\phi(C_{\text{goal}})\in\mathbb R^d$. The action space is

$$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$

where $(x,y,z)$ is end-effector position, $(R_x,R_y,R_z)$ is a ZYX-Euler rotation, $d_{ee}$ is finger separation, and $\gamma\in\{-1,1\}$ is a binary “is_this_last_action?” flag. A trajectory has length $T\approx 25$–$N=2048$0.

The learning objective is

$N=2048$1

where $N=2048$2 is a shape-distance metric such as Chamfer Distance or EMD. Success is defined by a threshold on the final-state distance, for example $N=2048$3. This formulation makes the system explicitly goal-conditioned and long-horizon: the policy must map a history of point clouds, the previous action, and the goal latent to the next pinch action so that many small manipulations accumulate into the desired shape.

The hardware stack couples a Franka Emika Panda arm with a custom two-finger gripper whose soft fingertips are asymmetric: one concave and one convex “bone” 3D-printed core cast in soft EcoFlex skin. The paper notes that this asymmetric soft fingertip reduces harsh edges and requires full $N=2048$4 around-axis rotation in $N=2048$5. The dataset comprises 20 kinesthetically taught demonstrations of bowls with diameters in $N=2048$6 cm? No such interpolation is reported; the exact data are bowls of diameter $N=2048$7 and heights $N=2048$8, with each trajectory length in $N=2048$9. These are augmented by $C_{\text{goal}}\in\mathbb R^{N\times 3}$0 Z-axis rotations to yield 3,600 total trajectories.

3. Guided diffusion policy, latent geometry, and action filtering

PinchBot’s policy model builds on “Diffusion Policy” by Chi et al. 2023 and combines denoising-based action generation with learned geometric representations and explicit action filtering (Bartsch et al., 23 Jul 2025). The predicted variable is the next horizon of actions, $C_{\text{goal}}\in\mathbb R^{N\times 3}$1, with diffusion horizon $C_{\text{goal}}\in\mathbb R^{N\times 3}$2; at test time, the system executes four actions and then replans. A sub-goal policy with two sub-goals at steps of 8 is also evaluated.

The forward and reverse diffusion processes are specified as

$C_{\text{goal}}\in\mathbb R^{N\times 3}$3

and

$C_{\text{goal}}\in\mathbb R^{N\times 3}$4

with the simplified score-matching loss

$C_{\text{goal}}\in\mathbb R^{N\times 3}$5

Classifier-free guidance is optionally introduced through

$C_{\text{goal}}\in\mathbb R^{N\times 3}$6

which steers samples toward the goal latent.

The geometric encoder is pre-trained rather than learned from scratch in the low-data regime. Two backbones are compared: PointBERT, a Transformer pre-trained on ShapeNet via masked point modeling following Yu et al. 2022, and DP3-PointNet, a lightweight PointNet variant from Ze et al. 2024. Each produces a 512-dimensional latent after a 3-layer MLP head. The diffusion model also predicts a continuous progress variable $C_{\text{goal}}\in\mathbb R^{N\times 3}$7, interpreted as trajectory progress from start to end. The auxiliary loss is

$C_{\text{goal}}\in\mathbb R^{N\times 3}$8

and sampling terminates adaptively as $C_{\text{goal}}\in\mathbb R^{N\times 3}$9.

A further component is collision-constrained action projection. At each replanning step, the system fits a circle in the $g=\phi(C_{\text{goal}})\in\mathbb R^d$0 plane such that at least 95% of clay points project inside it. If a candidate action falls inside the fitted circle, its $g=\phi(C_{\text{goal}})\in\mathbb R^d$1 coordinates are projected to the boundary while keeping $g=\phi(C_{\text{goal}})\in\mathbb R^d$2, rotations, and $g=\phi(C_{\text{goal}})\in\mathbb R^d$3 unchanged. This is designed to reduce catastrophic deformations such as holes.

Reported results identify PointBERT with continuous progress guidance as the best-performing variant. For the 8 cm goal, the final metrics are $g=\phi(C_{\text{goal}})\in\mathbb R^d$4, $g=\phi(C_{\text{goal}})\in\mathbb R^d$5, and diameter $g=\phi(C_{\text{goal}})\in\mathbb R^d$6; for the 10 cm goal, $g=\phi(C_{\text{goal}})\in\mathbb R^d$7, $g=\phi(C_{\text{goal}})\in\mathbb R^d$8, and $g=\phi(C_{\text{goal}})\in\mathbb R^d$9; for the 12 cm goal, $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$0, $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$1, and $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$2. The ablations are particularly informative. On the 10 cm goal with PointBERT binary prediction, removing pre-training yields $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$3 mm, removing collision projection yields $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$4, and replacing diffusion with regression degrades performance sharply to $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$5 and $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$6. The discussion attributes strong performance primarily to pre-trained embeddings, progress prediction, and collision projection, while noting limitations in demonstration scale, shape diversity, and topology generalization.

4. Pinch end-effectors for linear parallel motion and adaptive grasping

A second major line of work uses pinch-based design at the mechanism level. The SP-Diff gripper employs a modular symmetric dual-finger configuration in which each finger combines a Semi-Peaucellier five-link linear actuator with a re-configured double-parallelogram mechanism for orientation stability, while a planetary differential inside a 60 mm-diameter by 30 mm-thick aluminium palm distributes torque to the two fingers (Ding et al., 18 Oct 2025). Its kinematic objective is strict linear-parallel grasping rather than the arc trajectories of conventional grippers.

For the SP mechanism, the fingertip displacement is given in closed form by

$$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$7

with small-angle linearization

$$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$8

The double-parallelogram arrangement maintains the distal phalanx exactly parallel to the base for all $$a_{\text{pinch}}=[x,y,z,R_x,R_y,R_z,d_{ee},\gamma],$$9. In the planetary differential, the torque balance is

$(x,y,z)$0

with ideal symmetric split $(x,y,z)$1, and if one finger meets an external resistance $(x,y,z)$2, the other finger takes $(x,y,z)$3 extra through the differential action. The prototype adopts $(x,y,z)$4 mm, $(x,y,z)$5 mm, $(x,y,z)$6 mm, $(x,y,z)$7 mm, identical 18 mm / 18 mm parallelogram links with 14 mm pivot spacing, an 18 T sun, 16 T planets, and a 50 T ring, with a 5:1 worm pre-stage and total actuation ratio of approximately 10:1. These values are reported to reduce the needed Z-axis touch-off by 30% relative to a 20 mm tilt-circle arc gripper. Experimental metrics include peak normal force per finger of 15 N in parallel mode, summed adaptive-grasp force of 25 N, a size range from 5 mm minimum diameter to 80 mm maximum envelope span, positional repeatability of $(x,y,z)$8 mm over 1000 cycles, and force repeatability of $(x,y,z)$9 N at $(R_x,R_y,R_z)$0.

The Hoecken-D Hand pursues the same combination of linear parallel pinching and self-adaptive grasping through a different mechanism: a modified Hoecken linkage plus a differential-spring structure (Guo et al., 15 Oct 2025). The straight-line constraint is

$(R_x,R_y,R_z)$1

with fingertip horizontal position

$(R_x,R_y,R_z)$2

Over the designed stroke $(R_x,R_y,R_z)$3, the deviation from a true straight line is reported as less than 0.5 mm. A contact-triggered transition to enveloping occurs when the differential spring torque satisfies

$(R_x,R_y,R_z)$4

The PLA prototype uses unit length $(R_x,R_y,R_z)$5 mm, $(R_x,R_y,R_z)$6 mm, $(R_x,R_y,R_z)$7 mm, torsional spring stiffness $(R_x,R_y,R_z)$8 N·mm/rad, and a pinch span of approximately 0–200 mm. Simulations report pinch force rising from about 10 N at $(R_x,R_y,R_z)$9 to about 70 N at $d_{ee}$0 when $d_{ee}$1. Experimentally, pure parallel pinch success exceeds 95% for rigid objects within span; thin-sheet success rises to 88% with mid-stroke enveloping; bottles and boxes in the 60–100 mm range achieve 90%; and overall stable grasps are reported across more than 90% of varied geometries.

Taken together, these two grippers show that recent pinch-end-effector design emphasizes straight-line guidance, underactuation, and contact-triggered morphological adaptation. A plausible implication is that some of the adaptability traditionally assigned to sensing and high-level control is being reallocated into linkage geometry and differential transmission.

5. Tactile sensing, stable pinching, and few-shot imitation

Another PinchBot-associated strand concerns tactilely closed-loop pinch grasping. In “A Robust Controller for Stable 3D Pinching using Tactile Sensing,” the grasp is stabilized by rolling the fingertips on the contact surface and applying a desired grasping force to reach an equilibrium state (Psomopoulou et al., 2021). The formulation uses non-penetration and rolling constraints, decomposes fingertip rolling into the orthogonal contact-surface angles $d_{ee}$2 and $d_{ee}$3, and commands finger torques through a control law that combines joint damping, desired pinch force $d_{ee}$4, and rolling terms derived from contact-frame geometry. The closed-loop system is shown to converge to equilibria in which all velocities tend to zero, the line between fingertip centers becomes parallel to the line between contact points, and each contact force has magnitude $d_{ee}$5.

The tactile subsystem uses TacTip optical sensors with a Sony IMX219 internal camera at $d_{ee}$6 px and output images cropped to $d_{ee}$7 px. First contact is detected via

$d_{ee}$8

and local surface orientation is estimated by a deep CNN trained on 5,000 random contact poses over ranges $d_{ee}$9, roll $\gamma\in\{-1,1\}$0, and pitch $\gamma\in\{-1,1\}$1. Reported mean absolute errors remain below $\gamma\in\{-1,1\}$2 for orientations. Real-robot experiments on a Shadow Modular Grasper with TacTip-equipped fingers use a grasping force target of 10 N and show stable equilibrium on objects including a stack of Post-Its, an empty cardboard box, a plastic lemon, and a brain-shaped stress toy; external pushes induce transient velocity spikes followed by convergence to a new equilibrium.

A more data-driven tactile route appears in “Learning Fine Pinch-Grasp Skills using Tactile Sensing from A Few Real-world Demonstrations,” which uses imitation learning for fine bimanual pinch grasping from only five teleoperated demonstrations on dual Panda arms controlled through two Sigma7 devices (Mao et al., 2023). The tactile encoder is a convolutional autoencoder operating on $\gamma\in\{-1,1\}$3 grayscale tactile images, with four stride-2 convolutional layers producing a latent tensor $\gamma\in\{-1,1\}$4 and a symmetric transposed-convolution decoder trained by binary cross-entropy reconstruction loss. The policy state is

$\gamma\in\{-1,1\}$5

where $\gamma\in\{-1,1\}$6 is the tactile latent, $\gamma\in\{-1,1\}$7 is the proprioceptive vector, and $\gamma\in\{-1,1\}$8 is the tactile image difference from a no-contact reference. After tiling, these are concatenated into a $\gamma\in\{-1,1\}$9 input for a small BC convolutional stack. The imitation objective is

$T\approx 25$0

The learned policy implicitly performs active contact searching: as soon as the tactile difference exceeds a small threshold, it transitions from pre-grasp motion into press and roll actions, then maintains or re-establishes stable contact. Quantitatively, the reported pinch-grasp success rate is 95% over 20 trials with object pose variation, greater than 90% robustness under external pushes, 100% recovery after drop within workspace, and 90% success on ten unseen cylinders without retraining. Test loss converges to approximately 0.04 for the proposed BC method, compared with approximately 0.5 for frozen tactile encoding and approximately 1.0 for a proprioception-only FCN. Saliency analysis based on the method of Simonyan et al. shows dynamic redistribution of weight between tactile and proprioceptive modalities across pre-grasp, press, roll-lift, and stabilization phases.

These two works jointly show that tactile information can enter pinch manipulation at very different levels: analytic state estimation and Lyapunov-stable control in one case, latent encoding and behavior cloning in the other. In both, touch is not an auxiliary sensor channel; it is part of the core state representation.

6. Wearable assistance, pinched structures, and emerging directions

In the assistive robotics setting, SoftPINCH presents an EMG-driven soft wearable exoskeleton for thumb-index finger flexion and pinch-grasp assistance, and the detailed summary explicitly describes this embodiment as a “PinchBot” realization (Grønvall et al., 3 Jun 2026). The device is built around the SoFiE platform, with a single Pololu 250:1 micro-metal-gear motor housed on the upper arm, tendon-driven actuation routed through PTFE Bowden tubes, and a floating pulley that splits force approximately 50%/50% between thumb and index so that one digit can continue moving if the other is constrained. The motor can deliver on the order of 0.12 Nm continuous torque at the tendon; with pulley radius of approximately 5 mm, this corresponds to about 24 N of pinch force.

Surface EMG is recorded from three Delsys Trigno Avanti sensors at 2000 Hz over forearm muscle groups. Preprocessing includes 20–450 Hz fourth-order Butterworth band-pass filtering, 50 Hz notch filtering, Hampel outlier removal, a 250 ms RMS envelope with 25 ms step, z-score normalization, and segmentation into 9 s epochs. Three subject-independent decoders are evaluated: LSTM, CNN+LSTM, and CNN+LSTM with attention. The CNN+LSTM and CNN+LSTM-attention models both achieve 99.4% LOSO test accuracy, outperforming the standalone LSTM at 97.8%; the simpler CNN+LSTM is selected for real-time deployment because attention does not provide a significant improvement and inference time is 30% lower on the ESP32-based controller. The intent-to-command stage maps class probabilities to a desired grasp angle, the motor torque is generated by a proportional law $T\approx 25$1 with $T\approx 25$2, and magnetic fingertip sensing stops closure when the field crosses a grasp threshold of about 5 mT. Active assistance reduces muscular effort during loaded grasping across all tested loads, with a reported 92.6% reduction at 1 kg.

A structurally different use of pinching appears in the planar 3-degree-of-freedom manipulator based on pinched bistable tapes (Osele et al., 2021). Here, pinching is not a grasping action but the physical mechanism by which a revolute joint is formed. Two back-to-back tape-measure-like bistable tapes are locally flattened by a traveling pinching node, and the flattened band behaves like an Euler–Bernoulli beam. The manipulator is modeled as a planar PRP chain with generalized coordinates $T\approx 25$3, $T\approx 25$4, and $T\approx 25$5, forward kinematics

$T\approx 25$6

and a cable-angle relation

$T\approx 25$7

Pinching drops bending stiffness by more than an order of magnitude: in back-to-back tests, the unpinched pair requires up to approximately 0.65 N·m to reach $T\approx 25$8, whereas the pinched pair requires only approximately 0.055 N·m. The reported working envelope reaches up to 2 m with angular span $T\approx 25$9, positioning repeatability within one 3 in grid cell, payload of about 0.2 kg for link lengths around 1 m, total hardware mass of 535 g excluding off-the-shelf tape cans, and extension ratio greater than 20:1 when loaded in tension. The long-term goal is UAV integration.

Across the cited literature, future directions are correspondingly diverse. The pottery PinchBot paper proposes scaling demonstrations with synthetic or simulated data, incorporating learned dynamics models such as graph-networks, and steering diffusion policies through latent-space RL following Wagenmaker et al. 2025 (Bartsch et al., 23 Jul 2025). The SP-Diff gripper embeds force-sensor PCB pads, a 12 × 12 mm micro-vision window, and EtherCAT-based streaming of encoder, carrier-angle, and slider-position signals for digital twin frameworks (Ding et al., 18 Oct 2025). The tactile-stabilization work explicitly points toward in-hand manipulation and multi-finger scaling (Psomopoulou et al., 2021). What remains consistent across these trajectories is the view of pinching as a high-value primitive for robotics: it can be a shaping action, a grasp mode, a tactile control problem, a wearable assistance channel, or a structural actuation principle.