Tactile-Action Mixed Controller

• Tactile-action mixed controller is defined by its integrated, bidirectional loop where tactile feedback directly modulates subsequent actuation.
• It combines sensor dynamics with control laws, enabling applications from musical expression to robust contact-rich robot manipulation.
• The approach leverages predictive tactile dynamics and multi-timescale correction for enhanced task performance and safety.

to=arxiv_search 凤凰大参考 0 0 {"query":"\"tactile-action mixed controller\" OR tactile action mixed controller arXiv", "max_results": 10, "sort_by":"relevance"} to=arxiv_search 天天彩票与你同行 彩神争霸是不是ict 0 {"query":"cyclotactor tactile platform musical interaction (Jong, 2021)", "max_results": 5, "sort_by":"relevance"} to=arxiv_search 大发快三豹子 】【。】【”】【json {"query":"Tactile-VLA Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization", "max_results": 5, "sort_by":"relevance"} A tactile-action mixed controller is a control architecture in which tactile sensing and action are coupled within a single closed loop, so that tactile feedback is not merely observed after the fact but participates directly in generating, modulating, or refining the next action. In the strongest formulations, tactile output becomes a cause of subsequent tactile input, contact forces enter the control law explicitly, or predicted and observed tactile dynamics are fused with action generation at multiple timescales. Across musical interaction, haptic user interfaces, and contact-rich robotics, the term therefore denotes more than tactile augmentation: it denotes control in which touch and actuation are co-constitutive parts of the interaction state (Jong, 2021, Aydinoglu et al., 2019, Shirai et al., 2023, Bi et al., 23 Jul 2025, Zhang et al., 30 Jun 2026).

1. Core concept and distinguishing features

The defining property of a tactile-action mixed controller is that tactile information is not treated as a passive side channel. In the cyclotactor, tactile output “is not merely a response to input, but also a cause of subsequent input changes” (Jong, 2021). In contact-aware robotics, the controller is written directly as

$u(x,\lambda)=Kx+L\lambda,$

so measured or estimated contact force $\lambda$ is part of the feedback law itself (Aydinoglu et al., 2019). In recent visuo-tactile robot policies, the same idea appears in learned form: tactile feedback updates instructions at the semantic level, refines action chunks at the control level, or is predicted as a future interaction variable rather than concatenated as a static observation (Bi et al., 23 Jul 2025, Tian et al., 2 Jul 2026, Zhang et al., 30 Jun 2026, Zheng et al., 19 Mar 2026).

A second distinguishing feature is bidirectionality. Conventional tactile displays and many vision-tactile policies are one-way: input is sensed, then output is issued. Mixed controllers instead encode reciprocity. The finger, tool, gripper, or end-effector changes the tactile state; the tactile state modifies the next action; the new action changes contact again. This reciprocal structure is explicit in systems as different as a finger-worn musical electromagnet, a complementarity-based multi-contact controller, and a 60 Hz reflexive latent tactile controller (Jong, 2021, Aydinoglu et al., 2019, Zheng et al., 19 Mar 2026).

System	Mixed-controller mechanism	Reported role
Cyclotactor	Closed tactile input/output loop	Multiplexes nearness and rigidity
Contact-aware controller	$u=Kx+L\lambda$ in an LCS	Stable control across contact transitions
Tactile Tool Manipulation	Tactile pose estimator + MPC replanning	Recovers from unexpected contacts
VLA-Touch	Semantic tactile planning + interpolant controller	Improves planning and execution
UniTacVLA	Predicted and current tactile latents refine action	High-frequency residual correction
OmniVTA	Predicted vs. observed tactile mismatch drives RLTC	Closed-loop contact correction

This suggests that the phrase names a family resemblance rather than a single implementation. What unifies the family is the causal status of touch in control.

2. Embodied musical and audio-tactile interaction

The paper “The cyclotactor: towards a tactile platform for musical interaction” defines the concept in especially explicit form (Jong, 2021). The device is a finger-based tactile I/O controller for musical interaction built from an electromagnet, an infrared-reflection proximity sensor, and a combined permanent magnet / infrared reflector “keystone” attached to the finger with a Velcro strap. Its interaction plane is vertically adjustable, and a temperature sensor with temperature compensation is used to keep force output linear and consistent. The core loop is cyclical: finger motion changes magnetic output, that output alters finger position and vibration, and the changed finger state is sensed again. The controller is therefore mixed because the performer’s motion and the device’s force/vibration output form a single interactive loop.

The cyclotactor’s best-developed control strategy is the multiplexing of two degrees of freedom onto one tactile loop. The raw proximity signal contains a slow component reflecting intentional motion and a fast component reflecting actuation-induced vibration. The paper defines

$$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$

$$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$

and then derives

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$

using calibration curves measured at different nearness values. The normalization is based on “maximally rigid” and “maximally loose” fingers so that rigidity becomes orthogonal to nearness. In the musical demonstration, nearness controls center frequency and rigidity controls bandwidth for a single noise sound source, yielding a 2D sound space shaped both by finger position and by how stiffly the finger is held (Jong, 2021).

“The MATRIX: A Novel Controller for Musical Expression” presents a different tactile-action mixed controller based on a 12 by 12 array of spring-loaded rods, giving 144 independently sensed points across about 36 square inches (Overholt, 2020). The performer sculpts a 3D bed of rods with the hand; opto-electronics with quadrature encoding and 7-bit counters track rod motion; an FPGA implemented in VHDL transmits surface shape at 57.6 kbaud for an effective frame rate of about 30 Hz, which the paper describes as producing no perceptible delay. The system supports direct synthesis, signal processing, and algorithmic/gestural music. One rod can correspond to one harmonic in additive synthesis; in granular synthesis, rods can address grain length, level, pitch shift, and grain re-ordering; and in gestural music, activity level and excursion can drive drum-like behaviors (Overholt, 2020). Here the “mixed” character lies less in physical feedback from the device back into the hand than in the fusion of tactilely sculpted surface topology and real-time sound control.

The audio-tactile friction synthesizer of “Intuitive Control of Scraping and Rubbing Through Audio-tactile Synthesis” extends the mixed-controller idea into semantic action–object navigation (Aramaki et al., 2024). Sounds and tactile stimuli are organized by semantic labels for action and object, while both modalities are synthesized from signal descriptors including mean amplitude, standard deviation of amplitudes, and temporal distance between successive peaks. A pen-mounted vibrating actuator, used over a graphic tablet with headphones, adds a haptic channel to the original sound synthesizer. The perceptual result is not symmetry across modalities: in audition, impact distribution / temporal structure is the most influential parameter, whereas in touch the temporal parameter is not significant and amplitude variations matter more, with strong amplitudes associated with scratching and weak amplitudes with rubbing (Aramaki et al., 2024). A common misconception is therefore that a mixed controller should use identical cue structures for audio and touch. The reported evidence points in the opposite direction.

3. Formal feedback laws and switched tactile regulation

In robotics, one foundational formulation is “Contact-Aware Controller Design for Complementarity Systems” (Aydinoglu et al., 2019). The paper models contact-rich systems as linear complementarity systems with dynamics

$$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$

and then makes tactile/contact force feedback explicit through

$u(x,\lambda)=Kx+L\lambda.$

The resulting controller is “contact-aware” because it closes the loop on measured contact forces instead of using them only for guarded moves or hard-coded switching. Stability is certified with a piecewise quadratic Lyapunov function depending on both state and contact force,

$$V(x,\lambda)= \begin{bmatrix} x^T & \lambda^T \end{bmatrix} \begin{bmatrix} P & Q\ Q^T & R \end{bmatrix} \begin{bmatrix} x\ \lambda \end{bmatrix},$$

and the synthesis is non-combinatoric because the complementarity constraints encode contact modes implicitly rather than enumerating all $2^m$ possibilities. The reported examples include a cart-pole with soft walls, where the contact-aware controller succeeded 100% of the time over 100 random initial conditions versus 81% for LQR, and an acrobot with soft joint limits, where the contact-aware design succeeded 68% versus 29% for LQR (Aydinoglu et al., 2019). This directly contradicts the misconception that tactile-action mixed control is necessarily heuristic or purely empirical.

A more explicitly reflexive architecture appears in “Tactile-Driven Gentle Grasping for Human-Robot Collaborative Tasks” (Ford et al., 2023). The Pisa/IIT SoftHand is equipped with five miniaturized TacTip optical tactile sensors, one on each fingertip, with each sensor processed asynchronously on a dedicated Raspberry Pi 4 Model B. Contact is estimated by comparing the current tactile image to a stored no-contact reference using the SSIM-based deformation signal

$\lambda$0

and the controller uses the mean deformation

$\lambda$1

as its tactile state variable. The switching logic is two-state: if no fingertip satisfies $\lambda$2, the hand closes with a proportional position controller toward maximum closure; if any fingertip is in contact, the controller switches to a PI regulator targeting $\lambda$3. The asynchronous architecture permits a 286 Hz control loop although each tactile camera runs at 30 Hz. Experimentally, the system stabilizes within 1–3 s after first contact, gently grasps 43 objects of varying geometry and stiffness, and achieves 80% success in a human-to-robot handover task, or 87% when partial successes are included (Ford et al., 2023).

These systems establish two enduring control patterns. One is direct tactile-force feedback inside a mathematically structured controller. The other is state-dependent switching in which fast movement transitions to tactile regulation once contact is established. Both are canonical mixed-controller motifs.

4. Tactile state estimation, grasp geometry, and contact-aware replanning

“Tactile Tool Manipulation” extends the mixed-controller paradigm to tool-mediated interaction, where the robot manipulates an external object through an external tool while maintaining simultaneous contacts (Shirai et al., 2023). The open-loop layer is a nonlinear trajectory optimization over object, tool, and gripper orientations, solved with IPOPT under quasi-static equilibrium, friction constraints, a kinematic force cone, and state/input bounds. The tactile layer is built on GelSlim 3.0 sensors that observe slip and relative motion at the tool-gripper interface. A tactile “stiffness regression,” trained with AprilTag ground truth, estimates the relative rotation $\lambda$4 of the tool in the gripper; a nonlinear least-squares estimator then infers object orientation $\lambda$5, pivot point $\lambda$6, and object length parameter $\lambda$7. On top of the offline plan, an MPC layer replans using the tactile estimate as the current state and tracks only the object orientation $\lambda$8, because tracking $\lambda$9 caused slipping. The reported implementation uses a horizon of $u=Kx+L\lambda$0 and runs at about 2 Hz. The main significance is structural: tactile sensing anchors the object-state estimate, and the action planner continuously reselects feasible tool motion under disturbance (Shirai et al., 2023).

“A Robust Controller for Stable 3D Pinching using Tactile Sensing” treats tactile perception as local geometry estimation for grasp stabilization (Psomopoulou et al., 2021). Two active fingers of a Shadow Modular Grasper carry BRL TacTip-based optical tactile fingertips. A convolutional neural network regresses contact depth $u=Kx+L\lambda$1, roll $u=Kx+L\lambda$2, and pitch $u=Kx+L\lambda$3 from the tactile images; these outputs define the contact-frame tangential directions $u=Kx+L\lambda$4 and $u=Kx+L\lambda$5. The torque controller combines joint damping, a desired grasping force $u=Kx+L\lambda$6 along the line between fingertips, and tangential rolling torques: $u=Kx+L\lambda$7 The design is explicitly reactive and does not require knowledge of the full object model, trajectory planning, or system dynamics. Reported mean absolute errors for the tactile pose estimator are approximately $u=Kx+L\lambda$8 and $u=Kx+L\lambda$9 for depth and $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$0, $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$1, $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$2, and $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$3 for angular components across the two sensors; hardware experiments show stable grasps on objects including a stack of post-it notes, an empty cardboard box, a plastic lemon, and a brain-shaped stress toy, with recovery after external pushes at approximately $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$4 s and $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$5 s (Psomopoulou et al., 2021).

“High-Bandwidth Tactile-Reactive Control for Grasp Adjustment” moves still further toward local tactile geometry as the sole feedback source (Lee et al., 19 Sep 2025). Fingertip tactile sensors operate at 200 Hz, and the controller alternates between gripper-closing mode and grasp-adjustment mode. In adjustment mode, fingertip linear motion is constrained to the contact tangent plane through projected descent,

$$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$6

while fingertip orientation is corrected if the contact point leaves a valid tactile region. For two-finger antipodal grasping, the paper identifies a failure mode of ordinary projected gradient descent and proposes Cross-Finger Gradient Descent (CFGD),

$$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$7

$$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$8

which raises simulation convergence from 1%, 12%, and 2% under PGD to 100%, 99%, and 99% under CFGD on ellipsoid, superquadrics, and torus objects, respectively. Hardware experiments report refinement to a stable antipodal grasp in about 1–2 seconds, with lifting when the mean angle drops below $$\text{nearness}(t)=\text{avg}(\text{proximity}(t)),$$9 (Lee et al., 19 Sep 2025).

Taken together, these works show that tactile-action mixed control need not be limited to force magnitude. It can depend on inferred tool pose, local surface orientation, contact points, contact normals, and tangent-space stability gradients.

5. Vision-language-tactile control and semantic force grounding

In large-model robotics, “VLA-Touch: Enhancing Vision-Language-Action Models with Dual-Level Tactile Feedback” introduces a modular two-level tactile feedback system layered on top of Robot Diffusion Transformer (RDT-1B) without fine-tuning the base VLA (Bi et al., 23 Jul 2025). At the planning level, Octopi converts a sequence of 6 GelSight Mini tactile frames into language describing properties such as hardness and roughness, and GPT-4o updates the manipulation instruction as

$$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$0

At the control level, a tactile-conditioned interpolant/diffusion controller based on BRIDGeR refines the VLA action chunk: $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$1 The controller uses a compact tactile force representation rather than raw tactile images,

$$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$2

where $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$3 and $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$4 are planar force direction components and $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$5 is force magnitude derived from the GelSight Mini $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$6 marker array. RGB observations are embedded with DINOv2, and inference uses 10 diffusion steps at about 8 Hz on an RTX 4090. The reported gains over base RDT are +42% on Cup, +140% on Wipe, and +67% on Peel; relative to a residual controller, the interpolant controller improves by +67%, +100%, and +42% on the same tasks (Bi et al., 23 Jul 2025). A central claim is that semantic tactile abstraction is better for high-level reasoning than raw tactile images, while low-dimensional force summaries are more appropriate for control.

“Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization” adopts a deeper fusion strategy (Huang et al., 12 Jul 2025). Vision, language, and tactile signals are encoded as a unified prefix sequence,

$$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$7

and the tactile-aware action expert predicts both target position $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$8 and target contact force $$\text{vibration deviation}(t)=\text{proximity}(t)-\text{nearness}(t),$$9. Execution uses a position-dominant hybrid position-force controller: $$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$0

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$1

followed by a PID joint controller. Net external force is regulated by Cartesian position, while internal grasping force is regulated by gripper width. The reasoning-augmented variant, Tactile-VLA-CoT, periodically analyzes failure and issues corrective language such as “The force was too light. A stronger force is needed. Now trying with 5N.” The reported experiments show language-to-force grounding on USB and charger insertion tasks, object-dependent grasp force selection on tabletop grasping, and adaptive correction on an unseen blackboard wiping task where force increases from 3.5N to 6.7N; success on the out-of-domain blackboard task is 80% for Tactile-VLA-CoT while baselines are 0% (Huang et al., 12 Jul 2025).

These two systems illustrate two complementary interpretations of “mixed.” In one, tactile information is translated into semantic feedback for a planner and into low-dimensional force features for a controller. In the other, tactile tokens are deeply fused with vision and language so that the policy outputs both motion and force targets. Both reject the assumption that contact-rich control can be reduced to kinematics alone.

6. Predictive tactile dynamics and multi-timescale control

A major recent development is the shift from using current tactile observations to modeling tactile dynamics as a control variable in their own right. “VT-WAM: Visual-Tactile World Action Model for Contact-Rich Manipulation” is exemplary (Tian et al., 2 Jul 2026). The model has visual, tactile, and action experts; visual observations $$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$2 and tactile deformation observations $$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$3 are encoded into visual and tactile tokens, while language instruction and proprioception are injected via cross-attention. Training minimizes a joint flow-matching objective,

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$4

and, when contact guidance is enabled,

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$5

The Asymmetric Mixture-of-Transformers Attention uses a first-frame visual anchor for context, while the action expert attends to full tactile dynamics; AVTAG then biases action queries toward tactile keys during contact phases. Across six real-world tasks on a 7-DoF xArm7 with paired Xense tactile sensors, VT-WAM achieves a 71.67% average success rate, outperforming Fast-WAM by 26.67% and OmniVTLA by 35.84%, with 90% on wipe board, 85% on wipe vase, 70% on peel cucumber, 60% on insert plug, 70% on swipe card, and 55% on insert tube (Tian et al., 2 Jul 2026). The paper’s central distinction is between conditioning on touch and predicting tactile deformation dynamics during action generation.

“UniTacVLA: Unified Tactile Understanding and Prediction in Vision Language Action Models” organizes the same principle around a unified tactile prior (Zhang et al., 30 Jun 2026). Real tactile observations are encoded into tactile latents; tactile chain-of-thought reasoning produces state-aware semantics; a coarse-to-fine predictor estimates future tactile states; and a tactile-action mixed controller then outputs bounded residual corrections: $$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$6 The backbone produces a low-frequency action chunk, shown at 2 Hz in the figure description, while the controller runs at high frequency, shown as 30 Hz, using current and predicted tactile latents. On the USB ablation, success rises from 30% with tactile input alone to 36% with T-CoT, 44% with coarse tactile prediction, 52% with fine tactile prediction, and 62% with the full action-tactile mixed controller (Zhang et al., 30 Jun 2026).

“Master Micro Residual Correction with Adaptive Tactile Fusion and Force-Mixed Control for Contact-Rich Manipulation” addresses the same problem through explicit multi-rate decomposition (Li et al., 16 Mar 2026). A 10 Hz Master-Guidance Policy generates temporally consistent action chunks, a 60 Hz Micro-Residual Corrector uses TCP wrench feedback for residual compensation, and a 125 Hz force-mixed PBIC execution layer converts pose references into compliant commands. Contact confidence is estimated as

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$7

and the PBIC layer computes

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$8

$$\text{rigidity}=\text{normalized}\big(\text{avg}(|\text{vibration deviation}(t)|)\big)$$9

$$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$0

The reported overall score is 0.94, with 93% damage-free success rate in chip grasping, 80% in plug insertion, 87% in block assembly, and 97% in surface wiping (Li et al., 16 Mar 2026).

“OmniVTA: Visuo-Tactile World Modeling for Contact-Rich Robotic Manipulation” situates a similar slow-fast controller inside a large-scale dataset regime (Zheng et al., 19 Mar 2026). OmniViTac contains $$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$1 trajectories across 86 tasks and $$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$2 objects. The slow path includes a TactileVAE, a two-stream Visuo-Tactile World Model, and an Adaptive Visuo-Tactile Fusion Policy; the fast path is the Reflexive Latent Tactile Controller operating at 60 Hz. The key tactile differential representation is

$$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$3

where $$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$4 is current tactile feature and $$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$5 is predicted tactile feature. The RLTC trains on abnormal-contact recovery segments and uses predicted tactile as a target for regulation rather than merely as an input (Zheng et al., 19 Mar 2026).

A plausible implication is that tactile-action mixed control is converging on a common architecture: coarse action generation at low frequency, tactile prediction or semantic abstraction for anticipation, and high-frequency corrective control for contact events.

7. Recurring themes, misconceptions, and conceptual boundaries

One recurring misconception is that tactile-action mixed control means only “adding touch” to an existing controller. Several papers explicitly reject that interpretation. VT-WAM argues that existing visual-tactile policies usually feed tactile observations directly into action prediction but “rarely model tactile deformation dynamics during action generation,” and UniTacVLA similarly distinguishes a state-aware and dynamics-aware tactile prior from passive tactile tokens (Tian et al., 2 Jul 2026, Zhang et al., 30 Jun 2026). OmniVTA makes the same distinction by using predicted tactile minus observed tactile as a control signal rather than using touch only as an auxiliary feature (Zheng et al., 19 Mar 2026).

A second misconception is that tactile representations should remain in raw sensor space. In practice, the surveyed systems use task-specific abstractions: SSIM deformation scores in gentle grasping, CNN-estimated surface orientation in 3D pinching, marker-derived force summaries $$\dot x=Ax+Bu+D\lambda, \qquad 0\le \lambda \perp Ex+F\lambda+c \ge 0,$$6 in VLA-Touch, Octopi-generated language for high-level planning, and latent tactile priors in world-model controllers (Ford et al., 2023, Psomopoulou et al., 2021, Bi et al., 23 Jul 2025, Zhang et al., 30 Jun 2026). This suggests that “mixed” refers less to a particular sensor encoding than to how tactile information participates in the control loop.

A third misconception is that one sensory invariant should transfer unchanged across modalities. The friction-synthesis study reports that temporal variation / impact distribution dominates in audition, while amplitude variation dominates in touch (Aramaki et al., 2024). The result matters beyond audio-tactile synthesis: it implies that mixed controllers may need shared high-level semantics with modality-specific low-level mappings.

Finally, the literature separates tactile-action mixed control from both purely reactive reflexes and purely deliberative planning. Complementarity-based control and tactile MPC show that tactile feedback can be embedded in formally constrained optimization with Lyapunov or trajectory-feasibility structure (Aydinoglu et al., 2019, Shirai et al., 2023). Diffusion- and world-model-based systems show that the same principle now scales to semantic planning, action-chunk refinement, and predictive contact modeling (Bi et al., 23 Jul 2025, Tian et al., 2 Jul 2026, Li et al., 16 Mar 2026). The concept therefore spans a broad design space, but its central criterion remains stable: tactile information must alter action generation in a causally closed and task-relevant manner, rather than serving only as post hoc sensing.