Article · AI translation
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, adding a crucial sense of touch to the robot brain. Pretrained on nearly 5,000 hours of real tactile data, the model covers three mainstream sensing principles—magnetic tactile, vision-based tactile, and piezoresistive sensing. It converts different signals into a shared tactile representation that can be used by different dexterous hands and action models, giving robots the tactile perception needed for fine manipulation alongside their ability to see clearly and reason effectively.
At a time when general-purpose large models are advancing rapidly, the decision to release a tactile foundation model for closed-loop physical manipulation signals Ant Lingbo's continued investment in and commitment to an embodiment-native approach.
Previously, Ant Lingbo's general-purpose robot brain had already developed capabilities in high-precision spatial perception, interaction with highly dynamic environments, and generalizable dexterous manipulation, enabling robots to see more clearly, reason more effectively, and act more deftly. LingBot-Touch has now been integrated into LingBot-VLA, extending perception to the “fingertips” so robots can “feel more sensitively” and further improving the capabilities of the embodiment-native general-purpose robot brain.
26 real-robot tests: Task success rates increase by 14 to 26 percentage points
As general-purpose large models continue to improve, robots are becoming better at understanding complex tasks and reasoning through the steps needed to perform them. But in the physical world, after every action, a robot must perceive changes in objects and its surroundings, then adjust its next action. In many fine manipulation tasks, a robot may see where an object is and move its hand to the right place, yet still fail to complete the task. Once its fingers touch the object, vision alone has difficulty determining whether the applied force is appropriate or whether the object is slipping; tactile sensing is needed to supplement it.
Consider a dexterous hand unscrewing a bottle cap: depth information can help its fingers line up with the cap. Once it begins turning, the cap may have loosened, or the finger pads may be slipping along its side. The two situations look similar in an image but require different adjustments. At that point, the position of contact at the fingertips, the amount of force, and how both change become important cues for the action model. Fingers and the object often obscure this information, making it difficult to obtain visually.
To test whether tactile representations help with physical manipulation, Ant Lingbo conducted 26 real-robot tests using the Sharpa Wave and XHand1 dexterous hands, covering fine manipulation tasks 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 operation without tactile sensing.
For example, when T-Rex controlled 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 directly into the model. It reached 90% after the tactile representations from LingBot-Touch were integrated.
(Figure caption: After LingBot-Touch was integrated, all four action models achieved their respective highest success rates on real-robot tasks using the two dexterous hands.)
A unified tactile representation lets robots reuse their “sense of touch” across hardware
Installing tactile sensors on a robot does not mean it can use touch effectively. Commercial tactile sensors as a whole are still at a relatively early stage of development. Mainstream approaches, including vision-based tactile, magnetic tactile, and piezoresistive sensing, each have advantages and limitations, and multiple approaches may coexist for a long time to serve different cost requirements and use cases.
A more difficult problem is that data collected by these sensors cannot readily be used interchangeably. The same contact produces different raw signals under different sensing principles and contact-point layouts. If a model learns to interpret only one type, its existing data and tactile experience may be difficult to reuse after a sensor or dexterous hand is changed, requiring new data collection, training, and adaptation. As hardware evolves rapidly, this tight coupling causes repeated development and raises adaptation costs for hardware makers, robot integrators, and model developers.
The key to deploying tactile sensing at scale, therefore, is to make tactile understanding independent of the physical signals of a particular sensor, so data from different sources can be used together for training and accumulated tactile capabilities can be reused across sensors, dexterous hands, models, and tasks.
To this end, LingBot-Touch uses separate encoding inputs for each sensing principle to process different raw signals, then maps contact points with different layouts into a unified fingertip coordinate space. A shared model learns the location and force of contact and how they change over time, producing unified tactile tokens. This allows data collected from a person wearing tactile gloves and from teleoperated dexterous hands to be used together for training. Different action models can also use the trained model to extract contact features. The architecture also leaves room to incorporate new types of tactile sensors in the future.
(Figure caption: LingBot-Touch encodes magnetic tactile, vision-based tactile, and piezoresistive tactile signals as unified tactile tokens for use by different action models.)
The team further validated the transfer of tactile capabilities across hardware on XHand1. The magnetic tactile sensor model used by XHand1 was absent from the pretraining data; all the model's experience with magnetic tactile sensing came from data collected by people wearing gloves. After switching to this dexterous hand, the team kept LingBot-Touch's parameters unchanged and trained only the action model using XHand operation demonstrations.
LingBot-Touch was also integrated with four different action models in these real-robot tests and is not tied to Ant Lingbo's own models. It provides general-purpose perception between tactile sensors and action models, rather than a dedicated module for a particular piece of hardware or policy. For the industry, this means sensor and dexterous-hand makers can choose hardware according to cost and task requirements, while robot integrators can retain their own action policies, reducing the repeated investment in developing tactile capabilities anew whenever a hardware and algorithm combination changes.
To balance perceptual capability with on-device computing resources, Ant Lingbo offers the model in two sizes. All 26 real-robot tests were completed using the lightweight LingBot-Touch-Tiny, which has approximately 5.3 million parameters. When integrated with LingBot-VLA, π0.5, and T-Rex, it increased the parameter count by only approximately 0.13% to 0.16%.
The larger LingBot-Touch-Small showed potential for further improvement in perception tests: when pretraining data increased to nearly 5,000 hours, the model continued to benefit from the additional data, indicating that tactile pretraining still has room to scale and improve.
Staying with an embodiment-native approach while advancing alongside general-purpose large models
Commenting on recent changes in the technical approaches to large models for embodied intelligence, Ant Lingbo CEO Zhu Xing said: “Advances in large language models for the digital world, represented by GPT-6 Astra and Claude Fable 5.2, have indeed strengthened robots' ability to reason through complex tasks and perform physical manipulation. We are paying close attention to this and actively incorporating those advances. But experience on the industry's front lines tells us that, to truly achieve a closed loop for manipulation, the difficult problem of generalizing physical manipulation cannot be sidestepped. To a large extent, it still depends on breakthroughs in embodied foundation models themselves.”
In his view, general-purpose large models and embodied foundation models have a synergistic relationship closer to 1+1>2: the former strengthen understanding of and reasoning about tasks and environments, while the latter improve physical perception and closed-loop manipulation. Advances in each reinforce the other and together 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. He added that Ant Lingbo will stay with its embodiment-native approach, continue exploring new limits for generalizing physical manipulation, build a general-purpose robot brain for the industry that works across configurations, promote the alignment of more physical modalities—including vision, depth, touch, and even temperature—within a unified model, and continue to release work as open source to the community while pursuing commercialization.
LingBot-Touch's technical report, model, and code have now been released as open source. Developers can access and download them on Hugging Face, GitHub, and the ModelScope community.
This article was provided by Ant Lingbo and republished by QbitAI with authorization. The views expressed belong to the original author.