蚂蚁灵波开源触觉基础模型 CEO朱兴：LLM可吸收 但泛化仍靠具身模型

Summary: Ant Lingbo has open-sourced LingBot-Touch, a tactile foundation model that gives robots a critical sense of touch.

1. Data and coverage: The model is pretrained on nearly 5,000 hours of real tactile data, covering three mainstream sensor types—magnetic tactile, visuotactile, and piezoresistive—and converts different signals into a unified tactile representation for use across dexterous hands and action models.

2. Performance data: The team ran 26 real-robot tests on two dexterous hands, Sharpa Wave and XHand1, covering fine operations such as turning book pages, unscrewing bottle caps, and recognizing mahjong tiles. With the model integrated, four action models—LingBot-VLA, π0.5, Diffusion Policy, and T-Rex—improved average task success rates by about 14 to 26 percentage points. In one case, T-Rex controlling Sharpa Wave to unscrew a bottle cap achieved 56.25% success without touch, 62.5% with raw signals directly input, and 90% after integrating the tactile representation.

3. Practical value: All 26 tests used the lightweight Tiny version with about 5.3 million parameters. Integration increased model parameter count by only about 0.13% to 0.16%, making it friendly to edge computing. A larger Small version is also available, and the model is open source, so changing hardware does not require retraining the tactile component.

Summary: Open-sourcing its self-developed tactile model is a dual play by Ant Lingbo in technical branding and ecosystem positioning.

1. Brand positioning: It adheres to an embodied-native approach and aims to build a general-purpose robotic brain across embodiments, aligning vision, depth, touch, and even temperature and other physical modalities within a unified model. CEO Zhu Xing stresses that the quality of modality input largely determines the quality of action output—no good input, no good output.

2. Ecosystem and channels: The model is not tied to its own action models. It has been integrated with four action models—LingBot-VLA, π0.5, Diffusion Policy, and T-Rex—and works with dexterous hands such as Sharpa Wave and XHand1. While advancing commercialization, it continues to open-source to the community, using open source to attract developers and hardware partners.

3. Product tiers: It offers Tiny (about 5.3 million parameters) and Small capacities, balancing perception capability and edge computing for different cost and task scenarios. It has a clear-eyed view that multiple commercial tactile sensor routes will coexist for a long time and does not bet on a single hardware path.

4. Credibility building: It supports brand credibility with verifiable data such as 26 real-robot tests, success-rate gains of 14 to 26 percentage points, and bottle-cap unscrewing improving from 56.25% to 90%.

Summary: This is an open-source opportunity that sellers can leverage, while the article also gives clear risk warnings.

1. Opportunity: The model is open source and open. Sensor and dexterous-hand makers can choose hardware based on cost and task requirements. Full-system teams can keep their own action policies, reducing the repeated investment of redeveloping tactile capabilities every time a hardware and algorithm combination changes.

2. Growing market: Demand is clear in fine manipulation scenarios for embodied intelligence, such as unscrewing bottle caps, turning book pages, and recognizing mahjong tiles. Tactile sensors as a whole are still at an early stage, with visuotactile, magnetic tactile, and piezoresistive routes coexisting for the long term, creating sustained demand for supporting solutions.

3. What to learn and ways to cooperate: A unified tactile representation lets tactile capabilities be reused across sensors, dexterous hands, models, and tasks. Data collected from people wearing tactile gloves and from teleoperated dexterous hands can be used together for training. The architecture reserves room for integrating new types of tactile sensors.

4. Risk warning: Raw signals are hard to generalize across different sensing principles and contact layouts. If a model learns only one type, existing data and experience are difficult to reuse after changing hardware, requiring new data collection, training, and adaptation. The lightweight version suits edge deployment, but its capability ceiling remains to be seen.

Summary: The direct value for production and design is freedom in selection, lower adaptation costs, and edge deployability.

1. Freedom in hardware selection: The model sits between tactile sensors and action models and provides general perception capability, rather than being a dedicated module for a particular hardware or policy. Manufacturers can choose sensors based on cost and task requirements, and full-system teams can keep their own action policies.

2. Cost reduction and reuse: When switching to the XHand1 dexterous hand, its magnetic tactile sensor model was not included in the pretraining data. The model's learned magnetic tactile experience came entirely from data collected with people wearing gloves. The team kept LingBot-Touch parameters unchanged and trained the action model only with XHand1 operation demonstrations, validating cross-hardware reuse.

3. Compute and deployment: All 26 real-robot tests used the lightweight Tiny version with about 5.3 million parameters. After integration with LingBot-VLA, π0.5, and T-Rex, parameter count increased by only about 0.13% to 0.16%, facilitating edge deployment. The larger Small version can still benefit from new data as the dataset grows to nearly 5,000 hours.

4. Production implications: The three routes—visuotactile, magnetic tactile, and piezoresistive—each have strengths and limitations and will coexist for the long term for different costs and use cases. It is advisable to lay out solutions by scenario and use open source to reduce repeated development and adaptation investment.

Summary: Customer pain points center on tactile data not being interoperable and hardware changes forcing rework; the solution is a unified tactile representation.

1. Customer pain points: The same contact produces different raw signals under different sensing principles and contact layouts. If a model learns to interpret only one of them, existing data and tactile experience are hard to reuse after changing sensors or dexterous hands, requiring new collection, training, and adaptation. In fine manipulation, fingers and objects occlude each other, making it difficult to judge whether force is appropriate or whether an object is slipping based on vision alone.

2. Solution: Separate encoding entrances are set up by sensing principle to process different raw signals first. Contacts from different layouts are then mapped into a unified fingertip coordinate space, where a shared model learns contact location, force, and their changes over time and outputs a unified tactile token.

3. Delivery form: It offers Tiny and Small capacities, balancing perception capability and edge computing. It is not tied to self-developed action models and has been integrated with four action models. The architecture leaves room for integrating new types of tactile sensors in the future.

4. Trend judgment: The key to scaling tactile deployment is freeing tactile understanding from dependence on specific sensor physical signals and enabling reuse across sensors, dexterous hands, models, and tasks.

Summary: Although the article has no platform player as its subject, LingBot-Touch is effectively building an intermediate platform at the tactile layer.

1. Demand side: With rapid hardware iteration, adaptation costs and duplicate development among sensor makers, full-system teams, and model developers are high. The industry needs a layer of general perception capability to decouple them.

2. Platform approach: It is positioned as general perception capability between tactile sensors and action models, not tied to Ant Lingbo's own models. It has been integrated with four action models—LingBot-VLA, π0.5, Diffusion Policy, and T-Rex—demonstrating an open-access posture.

3. Ecosystem recruitment and operation: It sets up separate encoding entrances by sensing principle and reserves expansion room for future new types of tactile sensors, making it easy for hardware partners to continue integrating. It continues to open-source to the community while advancing commercialization. It backs its claims with 26 real-robot tests, success-rate gains of 14 to 26 percentage points, and 90% on bottle-cap unscrewing, and lowers the integration barrier with a lightweight version that adds only about 0.13% to 0.16% parameters.

4. Risk avoidance: It judges that multiple technology routes for commercial tactile sensors may coexist for the long term and does not bet on a single hardware path. It takes a posture of actively absorbing rather than opposing general large models, while making clear that generalization in physical manipulation still requires breakthroughs in embodied foundation models themselves.

Summary: The article presents new developments, new questions, and route judgments in the tactile path of embodied intelligence.

1. New developments: The tactile foundation model is pretrained on nearly 5,000 hours of real tactile data, covering three sensing principles—magnetic tactile, visuotactile, and piezoresistive—and outputs a unified tactile token through a unified fingertip coordinate space. In 26 real-robot tests integrated with four action models, average task success rates improved by about 14 to 26 percentage points. After data grows to nearly 5,000 hours, the Small version still benefits from new data, indicating room for scaling tactile pretraining.

2. New questions: Commercial tactile sensors as a whole are still at an early stage, raw signals are not interoperable, and rapid hardware iteration brings binding and duplicate development. Generalization in physical manipulation remains a hard problem that cannot be avoided.

3. Route judgment: CEO Zhu Xing believes general large models and embodied foundation models are closer to a 1+1>2 synergy: the former enhances task and environment understanding and reasoning, while the latter improves physical perception and closed-loop manipulation. Progress in digital-world large language models can be actively absorbed, but generalization still depends on breakthroughs in embodied models. He also proposes that modality input quality determines action output quality and advocates aligning vision, depth, touch, and even temperature within a unified model.

4. Implications: The direction is a general-purpose robotic brain across embodiments, with commercialization and open source proceeding in parallel, using intermediate-layer perception capability to reduce industry adaptation costs.