NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Ant's Lingbot Open-Sources LingBot-Touch, Giving Robots a Shared Sense of Touch Trained on Nearly 5,000 Hours of Real Data

Ant Group's Lingbot released LingBot-Touch, an open tactile foundation model pretrained on nearly 5,000 hours of real contact data that unifies three sensor families into one representation — and lifts four different action models by 14–26 points in real-robot tests.

2026-11-18 · 1015 words · NeuroAI
Ant's Lingbot Open-Sources LingBot-Touch, Giving Robots a Shared Sense of Touch Trained on Nearly 5,000 Hours of Real Data

A robot can see a bottle cap perfectly and still fail to open it. The moment a fingertip contacts the cap, the decisive information — force applied, slippage starting, resistance changing — is hidden behind the finger itself. Vision cannot recover it. On November 10, Lingbot (蚂蚁灵波科技), the embodied-AI company under Ant Group (蚂蚁集团), open-sourced a model built to close exactly that gap: LingBot-Touch, a tactile foundation model pretrained on nearly 5,000 hours of real touch data — and one of the first attempts to make "feel" reusable across different robots, sensors, and policies.

The problem: touch is a fragmented language

Touch should be the easiest sense to justify. In fine manipulation — page turning, mahjong tile recognition, delicate assembly — contact states decide success, and camera views are occluded at the instant they matter most. Yet commercially available tactile sensors split across at least three incompatible principles: magnetic, vision-based (视触觉), and piezoresistive. Each differs in measurement physics, channel definitions, contact layouts, and sampling rates. Data collected on one sensor cannot train a model for another; every hardware iteration has traditionally meant recollecting data and retraining. That fragmentation has kept tactile data expensive at precisely the moment the embodied-AI industry is starved for it — real-robot data collection routinely costs hundreds of yuan per hour, around US$70 or more, and touch data is harder still.

LingBot-Touch's architecture is a direct answer. The model assigns each sensing principle its own encoding entry, processes raw signals separately, then maps all contact layouts into a standardized fingertip coordinate space. A shared representation network learns contact location, force magnitude, and how both evolve over time, outputting unified "tactile tokens" that any action model can consume. Because the encoders are separate, data collected by a human wearing a haptic glove and data from teleoperated dexterous hands can train the same model — and there is room to plug in new sensor types later.

What the real-robot tests show

Lingbot validated the model on 26 real manipulation tasks across two dexterous hands: Sharpa Wave and XHand1 (a fully direct-drive five-finger hand from Robot Era/星动纪元). Tasks ranged from page turning and bottle-cap twisting to distinguishing mahjong tiles. With LingBot-Touch plugged in, four different action models — LingBot-VLA, π0.5, Diffusion Policy, and T-Rex — averaged gains of 14 to 26 percentage points in task success over touch-free baselines.

One case makes the mechanism clear. With T-Rex controlling Sharpa Wave on cap twisting: no touch input, 56.25% success; raw sensor signals fed directly to the policy, 62.5%; the LingBot-Touch representation, 90%. The jump comes not from more data but from a better representation — the tactile model absorbs the work of interpreting raw contact signals, freeing the action policy to focus on acting.

Two further results stand out. Cross-hardware transfer worked: XHand1's magnetic sensor model was not in the pretraining data — all magnetic-touch experience came from glove-worn human collection — yet with the encoder parameters frozen and only XHand1 demonstrations for the policy, the system still delivered gains. And deployment is lightweight: all 26 tests ran on LingBot-Touch-Tiny, roughly 5.3 million parameters, adding only about 0.13–0.16% to host model size. Model weights, code, and the technical report are live on Hugging Face, GitHub, and ModelScope. Partner Pasini (帕西尼) contributed 1,136 hours of real magnetic tactile data and over 294,000 interaction records to pretraining.

The open-data bet, and an international echo

The release fits a widening pattern: as real-robot data becomes the binding constraint on embodied AI, open tactile datasets and models are emerging as the industry's public infrastructure. Lingbot's wider LingBot series has accumulated more than 33,000 GitHub stars and 2.2 million downloads in nine months of open-sourcing, and its LingBot-VLA general brain is adapted to 20 robots from 17 vendors. The same week, an international echo of the open-weights strategy arrived: Runway's Praxis-1 world action model, released under Apache 2.0 on November 11, gave robot developers a downloadable 27-billion-parameter action backbone. One team opened the fingertips, another opened the imagination layer — both are betting that the ecosystem around open models compounds faster than closed ones.

Lingbot CEO Zhu Xing (朱兴) framed the strategy plainly: general digital-world models keep improving task reasoning, but the hard part — generalizable physical operation — still depends on embodied foundation models themselves, and input quality decides output quality. The company plans to keep aligning more physical modalities, including depth and temperature, in one model.

Honest limitations

All performance numbers come from Lingbot's own technical report, self-tested on two specific dexterous hands and 26 selected tasks; success gains on other hardware, other task families, or in unstructured environments remain unverified. The 14–26-point improvements were measured against the team's own touch-free baselines, and baseline choice strongly influences such deltas. The cross-hardware claim rests on one demonstrated case (XHand1) with a frozen encoder — encouraging, but a sample size of one transfer. Tactile sensor maturity itself remains an industry bottleneck; the model cannot exceed the fidelity of the sensors feeding it. And real production-line evidence — long-run reliability, failure recovery, maintenance cost — does not yet exist for this release. Finally, the Praxis-1 comparison is directional rather than equivalent: the two models address different layers of the stack and are not directly benchmarked against each other.

What readers can do now

  • Reproduce the baseline. The full methodology — data hours, model weights, test protocol — is open; the fastest way to judge the claims is running LingBot-Touch-Tiny on your own contact-heavy task and comparing against your current policy.
  • Audit your sensor stack for representation risk. If your team trains touch policies tied to one sensor model, you are accumulating hardware-locked data; unified-representation models like this one are worth piloting before the next sensor purchase.
  • Watch who contributes data next. The architecture accepts new sensor types; the release's long-term value depends on whether other labs add their tactile datasets, turning a single-company model into a community resource.
  • Pair touch with your planning layer, not just your policy. The clearest gains appear where contact feedback changes mid-task decisions — gripping, twisting, insertion — so target integration at those decision points first.

Related coverage

Cite this article

NeuroAI (脑机网), "Ant's Lingbot Open-Sources LingBot-Touch, Giving Robots a Shared Sense of Touch Trained on Nearly 5,000 Hours of Real Data," 18 November 2026, https://neuroai.site/a/na-aigc-ant-lingbot-touch-open

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories