Compiled June 2026 · Focus: how touch and force are integrated into learned manipulation policies and VLAs — the model architecture (how tactile enters the policy), the sensor hardware, the 2026 venue trends (ICRA · ICLR · CVPR), and the open limitations.

This is the touch-centric companion to Dexterous Manipulation (§G tactile cluster), VLA Architectures §Category J (sensory-augmented input), and World Models (OmniVTA). For the full ICRA force/tactile cohort see ICRA 2026 Tactile/Force (92 papers).

Vision tells you where things are; touch tells you what is happening at the contact — force, slip, texture, and the moment of make/break contact. For insertion, tool use, fragile grasping, and in-hand manipulation, the decisive signal is at the fingertip and changes in milliseconds. VLAs inherited a vision-language prior with no tactile channel (π-series, GR00T, OpenVLA are touch-blind), so 2025–2026 produced a wave of work bolting touch onto policies. The central design questions are: (a) how does tactile enter the model, (b) what sensor produces it, (c) is the sensor needed at deployment, and — newly in mid-2026 — (d) does it transfer across sensors/embodiments, or is it welded to one rig? Question (d) is what the first tactile foundation policy, FTP-1, puts on the table.

flowchart TB
  Q{How is touch integrated?}
  Q --> A["A. Native tactile token<br/>fused into VLM/policy"]
  Q --> B["B. Modality expert + router"]
  Q --> C["C. Tactile world model<br/>+ reflexive control"]
  Q --> D["D. Distilled / sensor-free<br/>at deploy"]
  Q --> E["E. Frequency-aware<br/>F/T fusion"]
  Q --> F["F. Transient-contact /<br/>minimal sensing"]
  Q --> G["G. Cross-sensor foundation policy<br/>pretrain once, fine-tune per sensor"]

The field's central fault lines: the original split is sensor-free at deploy (D — FD-VLA, HapticVLA, distill touch away) vs sensor-in-the-loop (C/A — OmniVTA, native-token VLAs, servo on touch): sensor-free wins on cost/deployability, sensor-in-loop on reactive precision. Mid-2026 adds an orthogonal third axis — cross-sensor generality (G): is the policy welded to one sensor, or is the sensor interchangeable? The one-liner: FD-VLA removes the sensor, OmniVTA servos on the sensor, FTP-1 makes the sensor fungible. G is orthogonal because a foundation policy can itself be deployed sensor-in-loop or distilled sensor-free.

The 2026 message (ICRA especially): vision-based tactile sensors have splintered from a single GelSight clone into a hardware design space.

Tactile simulation & data (the enabling layer): 2026's clearest methodological shift is differentiable, physically-calibrated optical-tactile sim — DOT-Sim (2604.27367, Material Point Method + residual-image rendering, calibrates in minutes), differentiable path-tracing Sim2Real, ETac, TacFlex, ConTact (contrastive sim-to-real). On the data side, FreeTacMan (2506.01941) collects visuo-tactile demos via a robot-free wearable gripper (≈+50% over vision-only). The dominant practical lever remains data + sim, not the policy architecture.

The systems venue owns tactile hardware and deployment. Three shifts: (1) VBTS splintered into a design space (UV-encoded, Moiré, magnetic/EIT/acoustic/barometric, curved skins); (2) sim-to-real went differentiable + physically-calibrated (DOT-Sim, path-tracing, ETac, TacFlex, ConTact); (3) touch is being absorbed into learned policies/VLAs — including the headline force-without-a-sensor distillation (FD-VLA), hand-frame tactile anchoring for sub-mm mating (SaTA, 2510.14647, +30%), frequency-aware F/T fusion (ManipForce), transient-contact IMU sensing (TranTac), per-modality routing (Multi-Modal Consensus), and visuo-tactile world modeling (OmniVTA). Also: Symmetry-Aware Vision-Tactile (2602.13689). Trend: contact meets the robot — hardware, sim, and policy integration at volume.

The ML venue's tactile thread is representation-centric: learn a general, transferable tactile representation that survives sensor heterogeneity and captures dynamics. AnyTouch 2 (2602.09617) is the flagship — a general optical-tactile representation unifying object-level understanding with fine-grained, force-aware dynamic perception (ToucHD hierarchical dataset). DexMove brings a native-tactile (R-Tac marker) VLA with optical-flow contact features. Tactile also rides the broader sensory-augmented / unified-diffusion architectures here (it is the "J" modality in Review-VLA-Architecture). Trend: a tactile foundation-representation, sensor-agnostic and dynamics-aware — the missing "DINO for touch."

The vision venue frames touch through perception and deployment economics. HapticVLA (Skoltech) is the emblem: train an action model with safety-aware tactile rewards, then distill tactile knowledge into a vision-only VLA so it acts without a sensor at deploy (~87% real). The tactile/haptic/bimanual bucket sits alongside CVPR's egocentric and reasoning clusters. Trend: distill touch into vision; treat tactile as a training-time teacher, not a deploy-time requirement — the mirror image of ICRA's sensor-in-the-loop work.

Cross-venue arc: ICRA = hardware + sim + sensor-in-loop policies; ICLR = transferable tactile representation; CVPR = distill touch into vision, sensor-free deploy. The three are complementary: ICLR's representations feed ICRA's policies; CVPR's distillation is the deployment escape hatch from ICRA's sensor cost.

The arXiv frontier (mid-2026): the newest move closes the loop between ICLR's representation thread and ICRA's policy thread — a cross-sensor tactile foundation policy. FTP-1 (2606.13102, Tsinghua / Shanghai Qi Zhi / Yang Gao et al.) pretrains one policy on ~3,000 h across 21 sensors and fine-tunes per sensor by swapping only the encoder — the "OpenVLA/π0 moment" for touch. Not yet at a venue, but it is the first attempt to make a tactile policy a reusable checkpoint rather than a one-off.

1. Sensor heterogeneity — the universal tactile foundation is now being attempted, not yet settled. Encoders are largely sensor-specific (GelSight ≠ DIGIT ≠ magnetic). AnyTouch 2 / OmniVTLA push toward cross-sensor representations; FTP-1 (Jun 2026) is the first cross-sensor foundation policy, claiming +31% on unseen sensors by mapping image-/array-/state-based touch into a shared token space (MTTS). But it is a proof-of-concept — unseen-sensor eval is 3 tasks / 2 setups, and the gain partly reflects "tactile-aware vs. not" rather than purely "pretrained vs. from-scratch tactile." A robustly validated touch foundation model transferable across VBTS and force-based sensors is in progress, not established.
2. Sim-to-real gap. Tactile has a notoriously large sim-to-real gap (soft-body contact, optics). 2026's differentiable/calibrated sims (DOT-Sim, ETac) help but are mostly per-sensor and short-horizon.
3. Vision drowns out touch. Naive feature concatenation lets the high-dimensional vision stream dominate; routing (Multi-Modal Consensus) and frame-anchoring (SaTA) are partial fixes, not settled practice.
4. Sensor cost / durability / deploy. Fingertip sensors wear, add latency, and complicate fleets — the economic pressure behind the sensor-free distillation camp (FD-VLA, HapticVLA), which in turn cannot do true reactive correction.
5. Evaluation. Tactile work is largely real-robot-only with bespoke task suites — no standard tactile-VLA benchmark (LIBERO/SIMPLER are touch-blind), so cross-paper comparison is weak.
6. Language↔touch grounding is thin. Few systems connect tactile to language/instructions (VLA-Touch, OmniVTLA are early); "grasp it gently" / "until you feel a click" remains largely unsolved.
7. Reflex vs chunk mismatch. Chunk-level diffusion/VLA policies are too slow for contact reflexes; OmniVTA's 60 Hz loop is one answer, but reflexive correction is "largely missing from current visuo-tactile policies."

1. Need true reflexive contact correction (insertion, slip recovery)? → C: visuo-tactile world model + high-rate reflex (OmniVTA); or E if you have F/T (ManipForce).
2. Cost-sensitive fleet, can't put sensors on every robot? → D: distill at train, sensor-free at deploy (FD-VLA, HapticVLA).
3. Cheapest possible touch upgrade to a vision policy? → F: an IMU in the fingertip (TranTac).
4. Vision keeps overpowering touch? → B: per-modality experts + router (Multi-Modal Consensus), or A with frame-anchoring (SaTA).
5. Building a multi-sensor / transferable stack? → A with a general tactile representation (AnyTouch 2, OmniVTLA), or a pretrained cross-sensor foundation policy you fine-tune per sensor (FTP-1 — swap only the sensor encoder, reuse the shared tactile expert).
6. Blocked on data / sim? → FreeTacMan (wearable data) + a differentiable tactile sim (DOT-Sim/ETac) before touching the policy.

• Companion reviews: Dexterous Manipulation · VLA Architectures (Cat J) · World Models · VLM↔Action
• Venue cohorts: ICRA 2026 Tactile/Force (92) · ICRA 2026 survey · CVPR 2026 survey · ICLR 2026 survey
• Per-paper: FTP-1 · OmniVTA · FD-VLA · HapticVLA · AnyTouch 2

← Back to Home