Market

Robbyant open-sources LingBot-Touch, a tactile model trained on about 5,000 hours of touch data, with self-reported gains of 14 to 26 points across four policies

Ant Group’s Robbyant open-sources LingBot-Touch, trained on 4,983 hours of touch data. It reports 14–26 point gains across four policies on two robot hands.

By ·

Robbyant, Ant Group’s embodied-AI unit, open-sourced LingBot-Touch on November 10, a model that turns signals from three types of touch sensor into shared tactile tokens for robot policies. It was pretrained on about 4,983 hours of touch recordings, roughly two thirds from human operators wearing sensor gloves. Across 26 real-robot tasks on two dexterous hands, the company reports average success gains of about 14 to 26 percentage points over a no-touch baseline, spread over four policies and not for π0.5 alone.

The essentials

How the numbers were produced

The 26 tasks were split between a dual-arm Sharpa Wave setup (20) and an XHand1 setup (6), per Humanoids Daily. Each policy was post-trained without touch, with raw tactile signals and with frozen LingBot-Touch tokens; raw signals gained at most about half as much and sometimes lowered success. The ranges are therefore an average over four policies and two hands, not a per-policy figure. Interesting Engineering reports a bottle-cap test where success rose from 56.25% to 90%.

Limits the authors state

Only two hands were tested on real robots, capacitive sensors were absent from pretraining, and coverage centers on fingertips, so palm and side contact is underrepresented, per Humanoids Daily. The results come from the developer’s own evaluation. Humanoids Daily also discloses that Robbyant paid it for a separate sponsored thread, while saying its article was produced independently.

What is not established

Summary table

Summary table
PointWhat is stated as of November 7, 2026Source
ReleaseOpen-sourced November 10, 2026, Apache 2.0, code and weights on GitHub and Hugging FaceHumanoids Daily
Training dataAbout 4,983 hours of touch recordings, two thirds from sensor glovesHumanoids Daily
Real-robot tests26 tasks, two dexterous hands, four policiesHumanoids Daily
Reported gainAbout 14 to 26 points on average (13.9 to 24.8 in precise figures)Interesting Engineering
Model sizeTiny version of about 5.3 to 5.5 million parameters; reports differ slightlyInteresting Engineering

FAQ

Does LingBot-Touch improve π0.5 by 14 to 26 points?

Not by itself: the range covers average gains across four policies and two hands, π0.5 among them.

Is the data real or simulated?

The 4,983 hours are touch recordings, two thirds from human operators wearing sensor gloves and the rest from teleoperated robots.

Can anyone use it?

Code and weights are released under Apache 2.0 on GitHub and Hugging Face.

Method and limits

Checked on November 16, 2026. We read the Humanoids Daily page directly; Interesting Engineering and the Korea Newswire press release were seen through search excerpts. We did not read the technical report.

Sources

Spotted an error in this article? Report it · Our corrections policy

OrigoTech logo

New: the app iPhone & iPad

OrigoTech is now an iPhone and iPad app

Read the articles natively, with “New” markers on unread ones, local search and offline reading of what you have already loaded. Free, no account, iOS 17 or later; ads appear between articles.

Rekha icon: a house drawn in a single stroke

From the same author Self-promotion

An Eulerian path, drawn with your finger

Rekha, the iPhone game by OrigoTech's author: connect every dot in a single stroke without tracing the same link twice, across 500 drawings that are all solvable.

The OrigoTech newsletter

A sourced weekly digest of AI tool news.

Read next