Vision-based robot learning has crossed a threshold. Models like π₀.₅ and GR00T N1.5, pretrained on massive heterogeneous datasets, now serve as strong starting points for downstream manipulation — adapt the checkpoint, finetune on your task, deploy.

Tactile manipulation has had no equivalent.

Every tactile policy today is effectively bespoke: trained for one sensor (GelSight, Contactile, force-torque), one hand, one task suite. Switch the fingertip sensor or swap the robot platform, and you start from scratch. The reason is fundamental: tactile heterogeneity. A GelSight image, a Contactile array, and a 6-DoF wrench reading all encode touch — but in incompatible formats, resolutions, and morphologies.

In FTP-1: A Generalist Foundation Tactile Policy Across Tactile Sensors for Contact-Rich Manipulation — a collaboration among Sharpa, Tsinghua University, UC Berkeley, Shanghai Jiao Tong University, ETH Zurich, and others — researchers present the first generalist foundation tactile policy: pretrained on ~3,000 hours of data from 26 sources and 21 tactile sensors, transferable across embodiments, and surprisingly effective on sensors never seen during pretraining (+31% success-rate gain).

The Problem: Vision Has Foundation Models; Touch Does Not

Generalist vision-language-action (VLA) policies have shown that scale + heterogeneity → transferable skills. But contact-rich manipulation — insertion, force-controlled wiping, in-hand adjustment, cap twisting — depends on touch, not just sight.

Three barriers have kept tactile learning from scaling:

1. Sensor-specific silos

2. Existing tactile policies are tied to fixed hardware: one encoder architecture per sensor type, one embodiment, one observation format. Knowledge learned on GelSight does not transfer to Contactile; force-torque pretraining does not help vision-tactile finetuning.

3. Naïve tactile fusion hurts more than it helps

4. Simply injecting tactile tokens into a VLM backbone — as in Tactile-VLA — can degrade performance by interfering with pretrained vision-language knowledge. On Sharpa North long-horizon tasks, Tactile-VLA averages 35.8% success vs π₀.₅'s 45.3%. Touch without proper architecture is worse than no touch.

5. No shared pretraining corpus

6. Unlike vision (internet-scale images) or human video (EgoScale's 20k+ hours), tactile manipulation lacks a unified, large-scale dataset spanning diverse sensors, hands, and grippers — until now.

FTP-1 asks: Can one tactile policy absorb heterogeneous touch experience and transfer to sensors and robots beyond pretraining?

The Solution: Morphology-Aware Tokens + a Dedicated Tactile Expert

FTP-1 extends the multi-expert VLA architecture (built on π₀.₅) with two key innovations:

1. Morphology-Aware Tactile Token Space (MTTS)

2. A unified interface that maps any tactile input — image (GelSight), array (Contactile), or state (force/torque) — into 24 functional-area tokens representing semantically aligned hand regions: thumb tip, index tip, palm, wrist wrench, etc. A shared functional-area embedding tells the model where on the end-effector each token originates, regardless of which physical sensor produced it.

Parallel grippers map to thumb/index slots; dexterous hands use slots 0–14; wrist and finger F/T sensors use slots 15–20. Same semantics, different hardware.

2. Independent Tactile Transformer Expert (300M parameters)

Unlike adapter-based fusion, FTP-1 routes all tactile tokens through a dedicated tactile expert that the action head attends to — without back-propagating into the vision-language expert. This preserves VLM knowledge while learning reusable touch representations.

At finetune time on unseen sensors, only the sensor-specific encoder trains from scratch; the pretrained tactile expert, functional-area embeddings, and shared image-tactile Transformer chunk are reused.

The Data: FTP-1-Dataset — 3,000 Hours, 21 Sensors, 26 Sources

Pretraining scale matters. FTP-1-Dataset aggregates:

Critically, Sharpa contributed Sharpa North-FTP-1: 4,000 long-horizon dexterous demonstrations collected on Sharpa North with Sharpa DTC (Dynamic Tactile Array) sensors — one of the 21 sensors in the pretraining mixture. The paper acknowledges support from Sharpa Pte Ltd. for hardware, compute, and this dataset.

All annotations are standardized under MTTS, with language instructions rewritten via GPT-4o for linguistic diversity. This is the tactile analogue of what ImageNet did for vision — a shared representation layer over heterogeneous raw signals.

Putting It to the Test: 5 Institutions, 14 Tasks, Seen and Unseen Sensors

FTP-1 checkpoints were distributed to five independent institutions globally for downstream finetuning — a rigorous test of reproducibility and transfer.

Seen-sensor setups (in pretraining mixture)

Unseen-sensor setups (not in pretraining)

14 tasks span in-hand adjustment, force-controlled pressing, insertion/extraction, deformable-object interaction, and long-horizon bimanual manipulation.

The Results: +17.2% on Known Sensors, +31% on Unseen Ones

Simulation (UniVTAC) — FTP-1 achieves 66.7% average success (+17.5% over strongest baseline). On contact-dependent tasks excluding easy lifts: 59.5% vs 42.0% for architecture-only FTP-π₀.₅.

Real robots (seen sensors) — FTP-1 averages 62.5% (+17.2% over π₀.₅'s 45.3%):

Twist Cap and Wipe Dish are instructive: π₀.₅ pushes against caps without reactive force adjustment and loses dish contact during wiping. FTP-1 maintains stable pressure and slows insertion when tactile feedback signals misalignment.

Unseen sensors — the headline result:

+31.6 percentage points over the strongest baseline — with only the sensor encoder trained from scratch. On Insert Hanoi, FTP-1 exhibits reactive insertion control: slowing when misaligned. On Insert USB (100 demos only), it produces stable contact-aware motions where baselines shake and fail.

Ablation confirms this is transferable tactile knowledge, not just dataset proximity: a no-tactile-pretraining control (NTP-1) underperforms FTP-1 by +37.5% on unseen FlexivXense despite identical finetune architecture.

Why This Matters for Sharpa: From Hardware to Foundation Model

FTP-1 is not peripheral to Sharpa's stack — it is built on it.

Sharpa DTC as a pretraining sensor

Sharpa Wave's Dynamic Tactile Array (>1,000 pixels per fingertip, 6-DoF force/wrench via integrated API) is a first-class citizen in FTP-1's pretraining corpus. Policies finetuned on Sharpa DTC benefit from touch representations learned across 20 other sensors and thousands of hours of diverse contact.

Sharpa North as evaluation platform

Long-horizon real-robot benchmarks — Draw Balloon (deformable), Fix Hand (small-part assembly), Twist Cap (bimanual torque) — run on Sharpa North with dual 22-DoF Wave hands. These are the tasks where vision-only VLAs plateau and tactile foundation pretraining delivers the largest gains.

1:1 anthropomorphic action space

FTP-1's dexterous-hand data (30% of mixture) includes Sharpa North demonstrations. The 22-DoF joint-space control interface matches human hand kinematics — the same action space EgoScale uses for human-video pretraining and T-Rex uses for tactile-reactive control. Hardware, data, and model now align on one embodiment standard.

CraftNet VTLA validation

Sharpa's CraftNet argues that touch needs a dedicated high-frequency pathway (System 0) separate from vision-language reasoning. FTP-1 independently proves this at foundation-model scale: a 300M-parameter tactile expert outperforms adapter-based Tactile-VLA by +26.7 points on real-robot average — the architectural lesson CraftNet and FTP-1 converge on.

Sim-to-real and ecosystem

FTP-1 evaluators include UniVTAC simulation (visuo-tactile benchmark) and real Sharpa hardware. Wave ships with Isaac Sim assets in the NVIDIA Isaac GR00T Reference Humanoid — enabling teams to finetune FTP-1 in sim on Sharpa DTC parameters, then deploy on physical Wave hands.

Cross-sensor portability for OEMs

Robot OEMs integrating Sharpa Wave are not locked into one tactile encoder. FTP-1's MTTS means a policy pretrained on Sharpa DTC + GelSight + human data can finetune on Wave with minimal data — and the pretrained tactile expert transfers even when downstream users add unseen supplementary sensors.

The Bigger Picture: Touch Joins the Foundation Model Era

For years, tactile robotics faced a chicken-and-egg problem: no shared model because no shared dataset; no shared dataset because no shared representation. FTP-1 breaks the cycle with:

1. MTTS — a universal tactile token language across 21 sensors
2. FTP-1-Dataset — 3,000 hours of heterogeneous touch experience
3. Tactile Expert — reusable pretrained representations, not per-sensor scratch training
4. Open release — pretrained models, datasets, and training code at ftp1-policy.github.io

Together with SaTA (spatial grounding), Tacmap (tactile sim-to-real), T-Rex (high-frequency tactile reactivity), and CraftNet (hierarchical VTLA), FTP-1 completes the picture: tactile intelligence can be pretrained, shared, and transferred — the same way vision finally learned to scale.

For Sharpa, this is the software layer that makes Wave's hardware investment compound: every hour of tactile data collected on Sharpa platforms improves a foundation model that benefits the entire ecosystem.