Robot-hand and tactile-sensor makers can test a free, Apache-licensed touch encoder with existing control models instead of training one per sensor, cutting hardware integration costs.
Robbyant, the embodied AI company under Ant Group, released LingBot-Touch on November 10, an open model that turns signals from robot fingertip sensors into a common format that robot control software can use. TechNode reported that the company published the model weights and inference code, and that the model was trained on close to 5,000 hours of recordings from three kinds of touch sensors. Robbyant tested it on 26 real-world tasks using two different robot hands.
The claim that matters is portability. Touch sensors for robot hands work on different physical principles and produce data in different shapes, so a robot policy trained on one sensor usually cannot read another. LingBot-Touch is meant to sit between the sensor and the robot’s decision-making model, so that what one hand has learned about contact can be reused on another. According to The Paper, Robbyant says this lowers the cost of adapting AI models to new hardware.
The results so far are Robbyant’s own. The technical report, submitted to arXiv on November 9, has not been independently replicated, and the company’s figures rely on tests it ran itself.
Robbyant trained the model on three sensor families and tested it on two hands
The project page gives the training set as 4,983 hours, 464,431 recorded episodes and about 1.33 billion frames. Piezoresistive sensors, which measure pressure through changes in electrical resistance, account for 2,197 hours. Magnetic sensors contribute 1,136 hours and visuotactile sensors, which use a camera to watch a soft skin deform, 1,650 hours. The data came from people wearing tactile gloves as well as from robot hands.
Testing covered 20 tasks on the Sharpa Wave hand and 6 on the XHand, each run 20 times, plus 12 tasks in simulation. IT Home reported that the tasks included turning book pages, twisting bottle caps and identifying mahjong tiles by touch.
Robbyant connected the touch model to four existing robot control models: its own LingBot-VLA, π0.5, Diffusion Policy and T-Rex. Across the 26 tests, average task success rose by about 14 to 26 percentage points compared with the same models running without touch input, according to IT Home and NetEase. In one bottle-cap test, a T-Rex model driving the Sharpa Wave hand succeeded 56.25% of the time without touch and 90% with LingBot-Touch features. IT Home reported that feeding the same model raw sensor readings lifted the rate only to 62.5%, which is Robbyant’s argument for a learned representation over raw data.
A small encoder that leaves the robot’s main model untouched
The release comes in two sizes, LingBot-Touch-Tiny and LingBot-Touch-Small. All 26 real-world tests used the Tiny version, which the project page puts at 5.46 million parameters. NetEase reported that attaching it adds only about 0.13% to 0.16% to the size of the action model it feeds, and The Paper said that suits the limited computers carried on robots. Robbyant’s project page shows performance still improving as tactile pretraining data grows toward 5,000 hours, which suggests more data could help.
The cross-hardware test is the most telling. IT Home reported that the XHand uses a magnetic sensor model absent from the pretraining data; the model’s magnetic experience came from glove recordings. The team kept LingBot-Touch frozen and trained only the action model on XHand demonstrations.
The open release is narrower than a full training kit. The GitHub repository describes itself as an inference release under the Apache 2.0 license, with the frozen encoder, feature-extraction code and sample data for both hands. It does not include training code or the pretraining corpus, and downloading the model files requires an account with access to the repository. Both versions are listed in Robbyant’s Hugging Face collection, and Chinese media reported they are also on ModelScope.
Touch extends Robbyant’s push to be the software layer for other robot makers
LingBot-Touch adds a sense to the LingBot family Robbyant has been building since Ant Group set it up in late 2024. EastFrontier reported in August that Robbyant was seeking about RMB1.5 billion in its first outside funding round, pitching a software "robot brain" that works with many hardware makers. A shared touch encoder fits that strategy: it is most valuable to Robbyant if hand and sensor makers adopt it as a default.
Ant Group is also backing touch through investment. It led a strategic round in Daimon Robotics, which builds tactile sensing into its own robot manipulation model, EastFrontier reported in September, citing 36Kr. The Paper reported that Robbyant will keep to what it calls an embodied-native route, and that it sees general-purpose large models and robot foundation models as complementary.
The next useful signal is outside adoption: success rates reported by labs that did not build the model, and hand makers shipping LingBot-Touch support with their sensors.