Ant Group's Robbyant open-sources LingBot-Touch, a touch model that works across sensors and hands

Opening-sourcing LingBot-Touch

Robots are getting fingertips. They're still not very good at understanding what those fingertips are telling them. But that could soon change.

Ant Group's robotics arm Robbyant thinks part of the problem is the Tower of Babel in tactile hardware. Today it open-sourced LingBot-Touch, a tactile foundation model designed to read three very different kinds of touch sensors and translate them into a common language any robot brain can use.

The problem: every sensor speaks its own language

There's no standard robot fingertip.

Some tactile sensors are magnetic, some are optical (essentially a tiny camera looking at a deformable gel), and some are piezoresistive, measuring changes in electrical resistance under pressure. Each has different costs and tradeoffs, and Robbyant expects all three to coexist for a long time.

The catch is that the same physical touch produces completely different raw signals on each one. So a model trained on one sensor generally can't use another, and every time a team swaps hands or sensors, it re-collects data and retrains.

That's expensive.

And in an industry where dexterous hands are iterating every few months, it's a tax on everyone.

LingBot-Touch gives each sensor type its own tokenizer, maps every contact point into a shared fingertip coordinate space, then runs it all through one unified transformer. What comes out are "unified tactile tokens" describing contact location, force distribution and how both change over time. Those tokens can feed whatever policy is actually controlling the robot: a diffusion policy, or a full vision-language-action model.

One nice side effect: data from humans wearing tactile gloves and data from teleoperated robot hands can be trained together.

The results: 14-26% better

Robbyant pre-trained the model on nearly 5,000 hours of real-world tactile data, then tested it on 26 real-world tasks across two dexterous hands, the Sharpa Wave and the XHand. The tasks were the fiddly everyday stuff that can easily defeat robots:

• squeezing toothpaste onto a brush
• turning a key in a lock
• separating stacked cups
• pulling zippers
  

Four different action models got the tactile tokens: Robbyant's own LingBot-VLA, Physical Intelligence's π0.5, Diffusion Policy and T-Rex. Average task success rose by roughly 14 to 26 percentage points versus running with no touch at all.

The most telling example is unscrewing a bottle cap with T-Rex controlling a Sharpa Wave hand. No touch data achieved only a 56.25% success rate. But with raw tactile signals piped straight into the policy: that rose to 62.5%. Then with LingBot-Touch's processed representation it exploded to 90%.

In other words, just bolting sensors onto a hand barely helps, but interpreting what they say is where the gain is.

Robbyant also says contact localization error is 0.12 mm, versus 2.19 mm for FTP-1 using a platform-specific adapter.

Small enough for the control loop

The deployable version, LingBot-Touch-Tiny, is about 5.3 million parameters. That's tiny.

Bolted onto LingBot-VLA, π0.5 or T-Rex, it adds just 0.13% to 0.16% to total model size, which matters when touch has to be processed fast enough to keep a bottle cap from slipping.

There's also a larger version, which Robbyant says is still improving as pre-training data approaches 5,000 hours. That's good news: it means tactile pre-training hasn't hit a scaling wall yet.

The generalization claim

The most interesting result is on XHand.

Its magnetic tactile sensor never appeared in LingBot-Touch's pre-training data. All of the model's magnetic-sensing experience came from human glove data. Robbyant froze LingBot-Touch's parameters entirely and trained only the action model on XHand demonstrations, and it says the touch layer still delivered.

If that holds up outside Robbyant's lab, this is a very meaningful step: tactile understanding that transfers to new hardware without retraining.

Weights, code and the technical report are on GitHuband Hugging Face, so independent testing should come fairly quickly.

Touch is a crowded field

Touch is having a moment, you might say.

Just a couple of weeks ago, Zurich's Enact Intelligence raised $30 million on almost the opposite bet: skip fragile sensor skins entirely and measure force inside the motors. 1X has put high-resolution tactile sensing on NEO's fingertips, Genesis AI's Eno hands carry tactile sensors alongside an onboard camera, and researchers are building robot skin that triggers reflexes.

Robbyant's play is different from all of those.

It isn't building a sensor or a hand. It's trying to be the layer in between: the thing that lets hand makers pick whatever sensor fits their cost and task, while robotics teams keep their own policies.

And by open-sourcing it, Robbyant gets a shot at becoming the default.

For Ant Group, which says Robbyant's long-term goal is robots for elder care, medical assistance and household tasks, that's a smart way to get a lot of hands it doesn't own using software it does.

Whether everyone else wants a standard from Ant Group sitting between their sensors and their robot brains is a different question.

Sources

• Robbyant LingBot-Touch project page
• LingBot-Touch on GitHub
• LingBot-Touch on Hugging Face
• Robbyant press release, Nov. 10, 2026