Ant Group's Lingbo Technology Opens Tactile Foundation Model LingBot-Touch

On November 10, on the same day as the release of Claude Fable 5.2, Ant Lingbo Technology open-sourced the tactile foundation model LingBot-Touch. A general large model oriented towards the digital world and an embodied foundation model oriented towards physical operation closed-loop entered the public eye on the same day.

The relationship between these two technical routes is precisely the focus of discussion in the current embodied intelligence industry. In the view of Ant Lingbo, the general large model and the embodied foundation model are closer to a 1+1＞2 synergistic relationship: the former enhances understanding and reasoning of tasks and environments, while the latter complements physical perception and closed-loop operation capabilities; only by advancing both simultaneously can they jointly accelerate the maturity of the embodied industry.

The LingBot-Touch released this time is pre-trained based on nearly 5000 hours of real tactile data, covering three mainstream sensing principles: magnetic tactile, visual tactile, and piezoresistive. It can convert different signals into a single set of tactile representations for use by different dexterous hands and motion models, allowing robots to further complement the tactile perception required for fine manipulation while they “see clearly and think clearly.”

Previously, Ant Lingbo's general brain had already covered high-precision spatial perception, high-dynamic environment interaction, and generalizable dexterous manipulation; this time, LingBot-Touch has been connected to LingBot-VLA, extending perception capabilities to the “fingertips,” allowing robots to “touch more sensitively,” and taking the hand-eye coordination capability of the embodied native general brain to the next level.

26 real-machine tests, task success rate increased by 14 to 26 percentage points

As general large models continue to advance, robots' abilities to understand complex tasks and reason through operation steps are being enhanced. However, in the real physical world, for every step of action a robot performs, it must perceive changes in objects and the environment, then correct the next step of action. In many delicate operations, a robot can see where an object is and can extend its hand to the correct position, but it may not necessarily be able to successfully complete the task—after the finger touches the object, whether the force is appropriate or whether the object is slipping is difficult to judge by vision alone and requires tactile sensing to fill the gap.

Taking a dexterous hand unscrewing a bottle cap as an example, depth information can help the fingers align with the bottle cap; after starting to rotate, the bottle cap may already be loose, or the fingertip may be sliding along the cap wall. These two situations are similar in the image but require different action adjustments. The contact position of the fingertip, the magnitude of force and its changes are often obstructed by the fingers and the object, making them difficult to obtain via vision, yet they are precisely the important basis for judgment of the action model.

To verify the help of tactile representation for real operations, Ant Lingbo conducted 26 real-machine tests on two types of dexterous hands, Sharpa Wave and XHand1, covering delicate operations such as turning book pages, unscrewing bottle caps, and identifying mahjong tiles. After connecting LingBot-Touch, the average task success rate of four action models—LingBot-VLA, π0.5, Diffusion Policy, and T-Rex—increased by approximately 14 to 26 percentage points compared to when there was no tactile sensing. Taking T-Rex controlling Sharpa Wave to unscrew a bottle cap as an example, the task success rate without using tactile sensing was 56.25%, the success rate with direct input of raw sensor signals was 62.5%, and the success rate reached 90% after connecting LingBot-Touch tactile representation.

(Caption: After connecting to LingBot-Touch, the success rates of the four action models on real-machine tasks on the two dexterous hands all reached their respective highest levels)

Unify tactile representation, allowing “hand feel” to be reused across hardware, models, and tasks

Equipping a robot with tactile sensors does not mean it can use touch well. Currently, commercial tactile sensors as a whole are still in the early stages of development; mainstream routes such as visual touch, magnetic touch, and piezoresistive touch each have their own strengths and limitations. Facing different costs and usage scenarios, multiple technical routes may coexist for a long time. More troublesome is that the data collected by these sensors is difficult to generalize directly: the same contact will form different raw signals under different sensing principles and contact point layouts. Once a model learns to interpret only one type, after replacing the sensor or the dexterous hand, the existing data and tactile experience may be difficult to reuse, requiring re-collection, re-training, and re-adaptation.

Therefore, the key to the large-scale implementation of touch is to let tactile understanding break free from the dependence on specific sensor physical signals, enabling data from different sources to be used jointly for training, so that accumulated tactile capabilities can be reused across sensors, across dexterous hands, across models, and across tasks.

To this end, LingBot-Touch sets encoding entries according to sensing principles, first processing different raw signals, and then mapping contact points of different layouts to a unified fingertip coordinate space. A shared model learns the contact position, the force received, and their changes over time, outputting unified tactile Tokens. Thus, data collected by humans wearing tactile gloves and teleoperating dexterous hands can be used jointly for training, and different action models can directly call the trained model to extract contact features. This architecture also leaves expansion space for connecting new types of tactile sensors in the future.

(Figure caption: LingBot-Touch encodes magnetic touch, visual touch, and piezoresistive touch signals into unified tactile Tokens for use by different action models)

On XHand1, the team further verified the cross-hardware reuse of tactile capabilities: the magnetic tactile sensor model used by XHand1 was not in the pre-training data, and all magnetic tactile experience learned by the model came from data collected by humans wearing gloves. After switching to this dexterous hand, the team kept the LingBot-Touch parameters unchanged, and the action model could complete the transfer by only using the operation demonstrations of XHand for training. At the same time, LingBot-Touch was connected to four different action models in this real-machine test, and it is not bound to the self-developed model by Ant Lingbo; it provides general perception capabilities located between the tactile sensor and the action model, rather than a dedicated module for a specific hardware or a specific strategy. For the industry, sensor and dexterous hand manufacturers can choose hardware based on cost and task requirements, and the complete machine team can also retain their own action strategies, thereby reducing the repetitive investment of re-developing tactile capabilities every time a hardware and algorithm combination is replaced.

To balance perception capabilities and edge-side computing power, Ant Lingbo provided two capacity models: all 26 real-machine tests this time were completed using the lightweight version LingBot-Touch-Tiny with approximately 5.3 million parameters; after connecting to LingBot-VLA, π0.5, and T-Rex, the model parameter count only increased by about 0.13% to 0.16%. The larger capacity LingBot-Touch-Small showed potential for further improvement in perception tests—when the pre-training data increased to nearly 5000 hours, the model could still benefit from the added data, indicating that there is still room for continuous expansion and improvement in tactile pre-training.

Adhering to embodied native, advancing in synergy with general large models

Fable 5.2 and LingBot-Touch debuted on the same day, which in a sense is a microcosm of the current evolution of embodied intelligence technology routes. Regarding the impact of the accelerated evolution of general large models on the robot "brain" route, Zhu Xing, CEO of Ant Lingbo, stated: "The progress of digital world large language models represented by Claude Fable 5.2 and GPT-6 Astra has indeed made the complex task reasoning and physical operation capabilities of robots stronger, which is a point we are very focused on and are actively absorbing. However, practice on the industrial front tells us that to truly complete the operation loop, the hard bone of physical operation generalization still cannot be bypassed, and to a large extent still relies on breakthroughs in the embodied foundation models themselves."

In his view, general large models and embodied foundation models are closer to a 1+1>2 synergistic relationship: the former enhances understanding and reasoning of tasks and environments, while the latter improves physical perception and closed-loop operation capabilities; both break through individually and promote each other, jointly accelerating the maturity of the embodied intelligence industry.

“The quality of modal input largely determines the quality of action output; in other words, without good input, there is no good output,” said Zhu Xing. Ant Lingbo will persist in the embodied native route, continuously exploring new limits of physical operation generalization, creating a robot industry general brain across configurations, promoting the alignment of more physical modalities such as vision, depth, touch, and even temperature in a unified model, and continuously opening to the community as open source while advancing commercialization.

Currently, the technical report, model, and code of LingBot-Touch have been open-sourced, and developers can access and download them on Hugging Face, GitHub, and the ModelScope community.

LeFeng Network (Official Account: LeFeng Net) LeFeng Network