On November 10, Ant Lingbo Technology open-sourced the tactile foundation model LingBot-Touch
On November 10, Ant Lingbo Technology open-sourced the tactile foundation model LingBot-Touch, filling a critical gap in tactile perception for robot brains. The model was pre-trained on nearly 5,000 hours of real tactile data and covers three mainstream sensing principles: magnetic tactile sensing, vision-based tactile sensing, and piezoresistive sensing. It can convert different signals into a unified set of tactile representations for use by different dexterous hands and action models, enabling robots to further enhance the tactile perception required for precise manipulation while “seeing clearly and thinking intelligently.”
As general-purpose foundation models continue to advance rapidly, the launch of a tactile foundation model aimed at closing the loop in physical manipulation reflects Ant Lingbo Technology’s sustained investment in and conviction in an embodied-native approach.
Ant Lingbo Technology’s general-purpose brain had already covered high-precision spatial perception, high-dynamic environment interaction, and generalizable dexterous manipulation, enabling robots to see more clearly, think more intelligently, and act more skillfully. This time, LingBot-Touch has been integrated into LingBot-VLA, extending perception to the “fingertips” and enabling robots to “have a more sensitive touch,” further enhancing the capabilities of the embodied-native general-purpose brain.
26 Real-World Robot Tests: Task Success Rates Rise by 14 to 26 Percentage Points
As general-purpose foundation models continue to improve, robots are becoming better at understanding complex tasks and reasoning through manipulation steps. But in the real physical world, every time a robot performs an action, it must perceive changes in the object and environment before adjusting its next action. In many precise manipulation tasks, robots can clearly see where an object is and move their hands to the correct position, yet may still fail to complete the task. This is because once the fingers make contact with an object, whether the force is appropriate and whether the object is slipping are difficult to determine through vision alone and require supplemental tactile information.
Take a dexterous hand unscrewing a bottle cap as an example. Depth information can help the fingers align with the cap. Once turning begins, the cap may already have loosened, or the fingertip pad may be slipping along the cap’s sidewall. These two situations may look similar on screen but require different adjustments. At this point, the contact location, magnitude of the force, and changes in force all become important factors in the action model’s judgment, yet this information is often obscured by the fingers and object and is difficult to obtain through vision.
To verify how tactile representations help with real-world manipulation, Ant Lingbo Technology conducted 26 real-world robot tests using the Sharpa Wave and XHand1 dexterous hands, covering precise operations such as turning book pages, unscrewing bottle caps, and identifying mahjong tiles. After integrating LingBot-Touch, the average task success rates of four action models—LingBot-VLA, π0.5, Diffusion Policy, and T-Rex—rose by approximately 14 to 26 percentage points compared with their performance without tactile sensing.
For example, when T-Rex controlled the Sharpa Wave to unscrew a bottle cap, the task success rate was 56.25% without tactile sensing and 62.5% when raw sensor signals were fed in directly. After integrating LingBot-Touch’s tactile representations, the success rate reached 90%.
(Image caption: After integrating LingBot-Touch, all four action models achieved their respective highest task success rates in real-world robot tasks on the two dexterous hands.)
Unified Tactile Representations Enable Robots to Reuse Their “Sense of Touch” Across Hardware
Equipping robots with tactile sensors does not mean they can use touch effectively. At present, commercial tactile sensors are still collectively at a relatively early stage of development. Mainstream approaches such as vision-based tactile sensing, magnetic tactile sensing, and piezoresistive sensing each have their own advantages and limitations. Given different cost requirements and use cases, multiple technical approaches may coexist for a long time.
An even more thorny issue is that data collected by these sensors is difficult to use directly across applications. The same contact produces different raw signals under different sensing principles and contact-point layouts. If a model learns to interpret only one of them, existing data and tactile experience may be difficult to reuse after the sensor or dexterous hand is replaced, requiring data to be collected, models to be trained, and systems to be adapted again. Amid rapid hardware iteration, this coupling leads to redundant development and increases adaptation costs for hardware vendors, robot OEMs, and model developers.
Therefore, the key to deploying tactile sensing at scale is to make tactile understanding independent of the physical signals of any particular sensor, so that data from different sources can be used jointly for training and accumulated tactile capabilities can be reused across sensors, dexterous hands, models, and tasks.
To this end, LingBot-Touch provides separate encoding inputs for different sensing principles. It first processes the different raw signals, then maps contact points with different layouts into a unified fingertip coordinate space. A shared model learns contact location, force, and their changes over time, and outputs unified tactile Tokens. As a result, data collected by people wearing tactile gloves and by teleoperated dexterous hands can be used jointly for training, while different action models can use the trained model to extract contact features. This architecture also leaves room to incorporate new types of tactile sensors in the future.
(Caption: LingBot-Touch encodes magnetic, vision-based, and piezoresistive tactile signals into unified tactile Tokens for use by different action models.)
On XHand1, the team further validated the cross-hardware reuse of tactile capabilities. The model of magnetic tactile sensor used by XHand1 was not included in the pretraining data; all of the magnetic tactile experience learned by the model came from data collected by people wearing gloves. After switching to this dexterous hand, the team kept the LingBot-Touch parameters unchanged and trained the action model using only XHand operation demonstrations.
At the same time, LingBot-Touch was integrated with four different action models in this real-robot test and was not tied to Ant Lingbo’s proprietary models. It provides general perception capabilities between tactile sensors and action models, rather than a module dedicated to specific hardware or a particular policy. For industry, sensor and dexterous hand manufacturers can select hardware based on cost and task requirements, while robot OEMs can retain their own action policies. This reduces the repeated investment required to redevelop tactile capabilities whenever a hardware-and-algorithm combination changes.
To balance perception capabilities with on-device computing power, Ant Lingbo offers two model sizes. All 26 real-robot tests in this round were conducted using the lightweight LingBot-Touch-Tiny, with approximately 5.3 million parameters. After integration with LingBot-VLA, π0.5, and T-Rex, the model parameter count increased by only approximately 0.13% to 0.16%.
The larger LingBot-Touch-Small demonstrated further potential for improvement in perception testing. Even when the pretraining data increased to nearly 5,000 hours, the model continued to benefit from the additional data, indicating that tactile pretraining still has room for continued scaling and improvement.
Staying true to an embodied-native approach and advancing in synergy with general-purpose large models
In response to the recent evolution of embodied large-model technology, Ant Lingbo CEO Zhu Xing said, “Advances in digital-world large language models, represented by GPT-6 Astra and Claude Fable 5.2, have indeed strengthened robots’ ability to perform complex task reasoning and physical manipulation. We are paying close attention to this and are actively drawing on it. However, experience from the forefront of industry has shown us that, to truly make the manipulation loop work, the hard challenge of generalizing physical manipulation cannot be avoided and still depends to a large extent on breakthroughs by embodied foundation models themselves.”
In his view, general-purpose large models and embodied foundation models have a more symbiotic relationship of 1+1>2: the former enhances understanding and reasoning about tasks and environments, while the latter improves physical perception and closed-loop manipulation capabilities. Breakthroughs in both can promote each other and jointly accelerate the maturation of the embodied intelligence industry.
“The quality of modality inputs largely determines the quality of action outputs. In other words, without good inputs, there can be no good outputs,” Zhu Xing said. Ant Lingbo will continue to follow an embodied-native roadmap, continually explore new limits on the generalization of physical manipulation, and build a general-purpose robotics brain that works across embodiments. It will advance the alignment of more physical modalities—including vision, depth, touch, and even temperature—within a unified model, while continuing to open-source and share its work with the community as it advances commercialization.
Currently, the technical report, model, and code for LingBot-Touch have been open-sourced, and developers can access and download them on Hugging Face, GitHub, and the ModelScope community.
This article is provided by Ant Group LingBot. QbitAI is authorized to reprint it, and the views expressed belong to the original author.
