LingBot-Touch Open-Sourced: Unified Tactile Model Lifts Robot Success Rates 14-26%

LingBot-Touch: Open-Source Tactile Foundation Model for Dexterous Manipulation

Ant Group's LingBot series has open-sourced LingBot-Touch , a tactile foundation model designed to give general-purpose robot brains the missing sense of touch. The model is pre-trained on nearly 5,000 hours of real tactile data covering three mainstream sensing principles: magnetic tactile , visual tactile , and resistive . It converts disparate raw signals from different sensors into a single unified tactile representation, allowing accumulated "touch experience" to be reused across sensors, across dexterous hands, across action models, and across tasks .

Real-Robot Evaluation: 26 Tasks, Four Action Models

To validate the representation, Ant Group ran 26 real-robot tests on two dexterous hands — Sharpa Wave and XHand — covering fine-grained tasks such as page turning, bottle-cap unscrewing, and mahjong-tile identification. Four distinct action models were evaluated: LingBot-VLA , π0.5 , Diffusion Policy , and T-Rex . Across all models, adding LingBot-Touch tactile tokens increased average task success rates by 14 to 26 percentage points compared to using no tactile input.

Case Study: Bottle-Cap Unscrewing with T-Rex on Sharpa Wave

A concrete example illustrates the gain:

No tactile input: 56.25% success rate

Raw sensor signals fed directly: 62.5% success rate

LingBot-Touch unified representation: 90% success rate

This demonstrates that high-quality tactile information input can significantly boost robot action success rates .

Why Tactile Sensing Is Necessary Beyond Vision

In physical manipulation, vision alone often cannot determine whether contact forces are appropriate or if an object is slipping. For instance, when unscrewing a bottle cap, depth data helps align fingers, but once rotation starts the cap may be loose or the fingertip may be sliding along the cap wall — two visually similar situations that require different corrective actions. Contact location, force magnitude, and their temporal changes become critical decision signals for the action model, yet these are frequently occluded from cameras by the hand and object themselves.

Unified Tactile Representation Architecture

LingBot-Touch addresses the fragmentation of tactile sensor modalities (each with different raw signal formats and fingertip layouts) through a purpose-built architecture:

Sensor-specific encoder entries process raw signals from each sensing principle independently.

Unified fingertip coordinate space maps diverse fingertip layouts into a common spatial frame.

Shared backbone learns contact position, force, and their evolution over time.

Output: unified tactile tokens that any downstream action model can consume.

This design enables joint training on human teleoperation data (glove) and robot demonstration data , lets multiple action models share the same tactile features, and leaves room for future sensor types to be added without redesigning the whole pipeline.

Cross-Hardware Reuse Demonstrated on XHand

XHand employs a magnetic tactile sensor model that was not present in the pre-training dataset (magnetic data came only from human-glove collection). Ant Group froze LingBot-Touch parameters and trained only the action model on XHand demonstrations. The tactile perception transferred successfully, confirming that LingBot-Touch provides a general perception layer between sensor and policy , not a hardware- or policy-specific module. Sensor/hand vendors can choose hardware based on cost and task needs; whole-robot teams can keep their own action strategies, reducing redundant development when swapping hardware-algorithm combinations.

Two Model Capacities for Edge Deployment

LingBot-Touch-Tiny (~5.3M parameters) — used for all 26 real-robot tests. When plugged into LingBot-VLA, π0.5, and T-Rex, it adds only 0.13%–0.16% extra parameters.

LingBot-Touch-Small — larger capacity; in perception benchmarks it continues to benefit from scaling pre-training data up to 5,000 hours, indicating further headroom for tactile pre-training scaling.

Embodied-Native Philosophy and Roadmap

Ant Group frames this release as part of its "embodied-native" technical conviction: unlike digital-world multimodal models, robots operate in the physical world where information flows from sensors into the environment. Both spatial perception (depth) and contact perception (tactile, force, temperature) must be aligned in a unified model, because modality input quality largely determines action output quality . The team commits to continuing open-source contributions while advancing commercialization, aiming to push the generalization ceiling of physical manipulation across robot morphologies.

Resources

Project homepage: https://technology.robbyant.com/lingbot-touch Technical report: https://github.com/Robbyant/lingbot-touch/tree/main/paper GitHub repository: https://github.com/Robbyant/lingbot-touch Hugging Face collection: https://huggingface.co/collections/robbyant/lingbot-touch ModelScope community:

https://www.modelscope.cn/collections/Robbyant/LingBot-Touch