Written by | Wu Xianzhi Edited by | Wang Pan
Recently, as general-purpose large models continue to break through in task understanding and complex reasoning, the technological route for the robot "brain" has once again become a focal point of discussion in the embodied AI industry.
The latest Claude Fable 5.2 released by Anthropic demonstrated capabilities such as controlling robots to unscrew bottle caps and manipulate puzzles, achieving significant progress in robotic simulation evaluations. The potential for general models to extend into the physical world is being further validated.
This makes a specific question more concrete: When robots can already understand instructions and plan steps, how can they execute actions stably?
On November 10, Robbyant open-sourced the tactile foundation model LingBot-Touch, offering its own answer regarding contact perception. Pre-trained on nearly 5,000 hours of real-world tactile data, the model covers three sensing principles—magnetic tactile, visual tactile, and piezoresistive—and converts different signals into unified tactile representations for use by various dexterous hands and action models.
What it aims to supplement is the judgment after a robot contacts an object: Is the grip stable? Is slipping occurring? Should force be increased or released in the next step?
Training a General "Touch Sensation"
Touch sensation is typical tacit knowledge; Polanyi argued that humans know far more than they can articulate. This type of knowledge cannot be passed down through text but must be acquired through continuous practice. Similarly, for robots to master it, they need to accumulate experience.
The training method of LingBot-Touch illustrates the process of machines accumulating touch sensation.
In terms of learning duration, the pre-training of LingBot-Touch utilized 4,983 hours of real-world tactile data, comprising 464,000 interaction segments and approximately 1.33 billion frames. The actions covered gripping, rotating, sliding, wiping, assembly, and bimanual coordination.
The nearly 5,000 hours of data encompassed three tactile routes, including 2,197 hours of piezoresistive data collected via human-worn tactile gloves, 1,136 hours of magnetic tactile data, and 1,650 hours of visual tactile data from teleoperating Sharpa Wave dual-arm dexterous hand robots.
Two-thirds of the contact data was collected by humans wearing gloves, independent of the robot body itself, providing an alternative growth curve for scaling data accumulation.
The three types of sensors have different fingertip sizes, number of contact points, and layouts, requiring alignment on the model side. LingBot-Touch translates all different signals into tactile Tokens in the same language, labeling each Token with time, finger ID, and 3D position. During training, perturbations are applied to contact point density and position to prevent the model from memorizing the layout of any single sensor type.
Tactile signals are sparse. During the movement of a dexterous hand towards a target or while adjusting posture, sensor readings are close to zero, producing effective signals only at the moment of contact. For most of the time, there are no tactile signals. If trained using a "learn what you see" approach, the model may fail to learn anything effectively.
LingBot-Touch addresses this by masking partial contact points in a record, prompting the model to infer the likely contact in the masked region based on changes in the remaining points, and comparing this in feature space. It does not insist on reconstructing raw readings, thereby focusing training attention on key frames where contact occurs.
To adapt to edge-side models, the size cannot be too large. Once a cup in a robot's hand starts to slip, the time window for tactile computation vanishes instantly. Considering this, LingBot-Touch provides Tiny and Small versions. The lightweight Tiny version has approximately 5.3 million parameters. After integrating with LingBot-VLA, π0.5, and T-Rex, the total parameter count of the models increases by only 0.13%, 0.15%, and 0.16%, respectively.
Controlling the parameter count does not imply compromised performance.
Real-world tests were conducted using two dexterous hands: the high-degree-of-freedom Sharpa Wave and the low-degree-of-freedom XHand. Four action models were evaluated: Robbyant's own LingBot-VLA, and external models π0.5, Diffusion Policy, and T-Rex. Each combination was tested in three control groups: without tactile input, direct feeding of raw signals, and using LingBot-Touch extracted tactile Tokens.
Across 26 tasks completed by the two dexterous hands—including flipping pages, stacking cups, moving deformable cups, sorting by hardness/softness, and handing over boxes—the average success rate for all four action models was highest when using LingBot-Touch.
According to the technical report, in the scenario of zipping and unzipping, π0.5 achieved a 39% success rate on Sharpa Wave without tactile input. When directly fed raw signals, the success rate dropped to 15%. However, switching to LingBot-Touch tactile Tokens raised the success rate to 59%.
In terms of perception layer accuracy, LingBot-Touch improved the precision of locating the contact center. The average localization error when deployed on Sharpa Wave was 0.12 mm, whereas the control model FTP-1, even with a platform-specific adapter, had an error as high as 2.19 mm.
The Industry Gains a Standardized Solution
The difficulty in converting tactile experience accumulated across different hardware and tasks into reusable knowledge is currently the primary industrial pain point for robot tactile sensing.
Facing the parallel development of visual tactile, magnetic tactile, and piezoresistive sensing routes—with varying dimensions, contact point layouts, and sampling frequencies—every team developing action models must start from scratch relying on raw signals. Each route requires building its own collection pipeline, performing calibration and annotation independently. The resulting models serve only a single manufacturer's dexterous hand. Once the hardware iterates, the work must often start over.
Consequently, fine operations dependent on tactile perception have long been stuck at low levels of automation. Audi disclosed in early 2026 that the automation ratio for automotive wire harness production and assembly is still less than 10%. A single body wire harness contains up to 600 connectors. From picking up soft wire harnesses, inserting plugs into interfaces, to confirming latch engagement, almost every step requires simultaneous contact and adjustment. Such operations should be where tactile sensing plays the biggest role, yet costs remain high due to the inability to reuse experience.
In the past, tactile algorithms had to be adapted one-by-one to specific hardware models, representing a customized path. Customization problems require standardized solutions. LingBot-Touch took a different approach, adapting by sensing principle.
The encoding entry points are set according to principles, contact points are uniformly mapped to standardized fingertip coordinates, and layout perturbations are added during training. Thus, new models under the same principle can be directly integrated.
"Out-of-the-box" usability was verified on real robots. The magnetic tactile sensor model used by XHand did not appear in the pre-training data. Robbyant fixed the parameters of LingBot-Touch and trained the action strategy using only demonstrations from this hand. The success rates for all four action models improved significantly.
When LingBot-VLA moved deformable cups on XHand, reading sensor signals directly resulted in a 58% success rate. After integrating contact information extracted by LingBot-Touch, the success rate rapidly climbed to 90%. Switching to new tasks also does not require retraining the tactile model; simply teaching the action model with a small amount of demonstration data suffices.
Translating contact signals into the same language benefits the entire industrial chain.
Action model teams are direct beneficiaries. The same set of Tokens can be integrated into π0.5, Diffusion Policy, and T-Rex, eliminating the need to repeat investment in the tactile layer when switching models. Dexterous hand and sensor manufacturers, having handled device-level basics like calibration, noise reduction, and reading correction, can focus more R&D resources on precision and durability.
A general touch sensation gives OEMs freedom in component selection. Hybrid solutions—using visual tactile fingertips to distinguish fine contacts and piezoresistive arrays in the palm to sense large-area forces—become feasible from concept to reality. This applies equally to end-users. Factories and warehouses do not need to bind to a specific dexterous hand during selection; they can choose optimal hardware based on tasks, ensuring that touch sensation no longer constitutes a cost for changing components.
In the domestic context, general tactile models hold special value. China has the widest variety of dexterous hands and sensors globally, allowing the collection of contact data that single hardware types cannot cover. Diversity is a treasure trove but easily leads to fragmented experience. General tactile models allow different technological routes to move from distinct separation to complementary strengths, enabling data accumulated by every hardware unit to play a greater role.
Deploying a general touch sensation requires ecosystem support. Before releasing LingBot-Touch, the Robbyant embodied AI ecosystem had already begun to take shape. On the body side, partners include Leju Robotics and TiHu; on the scenario side, commercial tests are underway with companies like Guoda Pharmacy and Longsheng in retail sorting, logistics sorting, and industrial scenarios. Jianzhi Technology participated in co-building the data system.
All code and weights are open-source. External teams can download and use them immediately. Without collecting an hour of data or training a model for a day, existing robots gain touch sensation upon integration. While being open, real operational data continuously flows back for training. The more users there are, the more diverse the returned scenarios become. This is one reason why Robbyant's full-stack brain is difficult to replicate.
The Real World Has No Pause Button
Open source allows touch sensation to spill over, but what determines whether this path can maintain leadership is the judgment about the physical world itself.
In September, after the release of GPT-6 Astra, multiple research teams migrated it to robots for testing.
In RoboDojo's 42 manipulation tasks, without any specialized training, its comprehensive score surpassed all 40 public action policies on the leaderboard. The ability of general large models to understand tasks and plan steps seems directly transferable to robotics.
The same batch of tests also exposed shortcomings in general models. Tasks requiring high precision, dynamic control, and bimanual coordination proved challenging. When rotating objects in-hand, coordinating finger-to-finger contact was difficult. In a simulation walking test, the robot required 250 model calls for 30 seconds of walking, averaging nearly 40 seconds per call. The reason was that the simulation needed to pause and wait before computation completed.
The real world offers no opportunity for pauses. The layer following contact cannot be filled by general large models in the short term.
First, the difficulty of data acquisition differs completely. Data for language and vision models can be obtained from the internet, but data to cultivate touch sensation can only be collected in the physical world. Second, tactile signals change extremely fast, leaving a fleeting time window for inference. Models processing them must be sufficiently lightweight to embed in real-machine loops. Third, the zip/unzip case demonstrates that feeding raw signals directly to models can be counterproductive. Translating contact signals into usable information for models requires specialized data and professional training.
The layer not filled by general large models is precisely where Robbyant focuses. Robbyant chose an embodied-native approach, aiming to drive robots of different forms with a single brain, without needing to write separate programs for each hardware piece.
At the current stage, key capabilities required by robots are studied and validated separately. LingBot-Vision and LingBot-Depth 2.0 provide visual and spatial perception. LingBot-VLA 2.0 handles instruction understanding and action generation. LingBot-World 2.0 and LingBot-Video enable robots to perceive how the world changes over time. LingBot-VA 2.0 unifies world prediction and action generation. Post-contact processing is handled by LingBot-Touch.
Combined, these models cover the complete workflow of robot operation, closing the loop of perception, prediction, and execution.
Rhythmically, Robbyant first solidified vision, depth, and tactile modules before integrating them. LingBot-Touch has already fused with VLA. The next step is to align more modalities within the same model, gradually synthesizing a unified brain. Compared to general large models relying on internet data, Robbyant can establish a moat in the physical world.
Sometime in the future, in a warehouse, a robot will pick up a soft packaging bag it has never seen before. The judgment generated by the fingertip touching the object will already possess skilled touch sensation. The competition among dexterous hands will not cease, but new hardware and new tasks will no longer need to be learned from scratch.