A new arXiv paper, TouchAnything, asks the question every robotics lab on a tight budget has muttered into its coffee: can you read touch from a camera? Its answer is a dataset called EgoTouch — 208 manipulation tasks across 1,891 episodes, indoor and outdoor, each one a synchronised stack of head-mounted egocentric video, two wrist cameras, bimanual 3D hand pose, and continuous pressure maps from wearable tactile sensors. The framework on top, TouchAnything, takes the egocentric view as its primary input and predicts where and how hard the hands are pressing. Add the wrist views at inference and contact prediction improves by up to 5.0% Contact IoU and 6.1% Volumetric IoU.

Read that as a systems diagram, not a benchmark. The bottleneck in embodied AI has never really been compute or model size; it is physically grounded data. Vision scales — cameras are cheap, and egocentric footage is being collected by the warehouse-load, as the EgoVerse robot-learning guide from Labellerr lays out. Touch does not scale: high-quality tactile skins are expensive, fragile, and a nuisance to wire onto every gripper. TouchAnything’s move is the oldest trick in resilient architecture — pay the expensive cost once, then amortise it. Collect tactile supervision the hard way for a bounded dataset, train a model to predict it from pixels, and deploy on vision alone.

Which is also where the single point of failure hides. A vision-to-touch model is only as honest as the tactile dataset that supervised it — change the object materials, the lighting, the glove, and the inferred pressure can drift with no sensor to contradict it. PAZ’s archive tracks the lineage this sits in: the bimanual household-manipulation dataset (arXiv:2405.18860) and low-cost rigs like AhaRobot (arXiv:2503.10070) that tried to make the hardware cheap rather than optional. EgoTouch tries to make it optional. The field is converging on a handful of contact datasets as shared infrastructure — and infrastructure, as anyone who has drawn a real dependency graph knows, is the thing nobody audits until it breaks.

On a working bench this week, the practical reading is graceful degradation. The multi-view design means you use the sensors you have: egocentric-only when that is all there is, wrist cameras when the budget stretched. The 5–6% lift from wrist views is small, but it tells you where the marginal franc goes — not into a tactile array, but into one more cheap RGB camera mounted nearer the hands.

Atelier: Bring this to the fabrication cell. A robot seating a timber dovetail or troweling a surface needs to know contact and force, and PAZ’s archive has the robotic dovetail-and-finger-jointing work from ACADIA to show the joinery is real. Vision-inferred touch is a way to give a fabrication arm a sense of pressure without instrumenting the end-effector — a soft start toward force-aware robotic assembly on the building site, not just in the manipulation lab.

←TODAY: In 2026 tactile hardware still doesn’t scale; cameras do — EgoTouch infers contact from 1,891 video episodes.
→3012: In the Zurich-3012 city every surface reports its own contact state, because the sensor became the model, not the skin.
Fulcrum: You only have to instrument touch once; after that you can read it forever from cameras you already own.

Hack: This Hack teaches you to compute the Contact IoU the paper reports, so its 5% number stops being a press figure and becomes something you can measure on your own masks. Contact IoU is intersection-over-union on a thresholded pressure map — predicted contact against ground truth, over the hand surface. Compute it, render where the two disagree, and you have a failure map instead of a headline.

import numpy as np

def contact_iou(pred, true, thresh=0.5):
    p = pred >= thresh          # predicted contact
    t = true >= thresh          # measured contact
    inter = np.logical_and(p, t).sum()
    union = np.logical_or(p, t).sum()
    return inter / union if union else 1.0

The dataset, code, and benchmark are slated for public release. When they land, do the systems exercise, not the demo: clone the benchmark, run TouchAnything on your own egocentric clips, and go hunting for the failure mode you didn’t know you had — the object it reads as gripped when it isn’t. Draw that dependency graph this week, before a fabrication arm trusts it.

Sources & Further Reading

• Primary: arXiv — TouchAnything: Bimanual Tactile Estimation from Egocentric Video
• Reinforcing: Labellerr — EgoVerse dataset guide for robot learning