← All model news

Ant Group's Lingbo releases LingBot-Touch tactile foundation model: ~5,000 hours of data, one representation for four action models

According to a Machine Heart report on November 10, 2026, Ant Group's Lingbo released the tactile foundation model LingBot-Touch, unifying magnetic tactile, piezoresistive and vision-based tactile sensors with 4,983 hours of real tactile data; the report says the same pretrained representation improved real-robot success rates when plugged into Diffusion Policy, π0.5, LingBot-VLA and T-Rex.

Published Updated

What changed

According to a Machine Heart report on November 10, 2026, Ant Group's Lingbo released the tactile foundation model LingBot-Touch that day. The report opens with Anthropic's Fable 5.2 scoring 64.9% on robot manipulation demos and the RoboDojo benchmark, then focuses on the point that "a successful demo does not equal general physical manipulation ability," arguing that fine "feel" still needs embodied native models. The report says the team collected 4,983 hours of real tactile data for LingBot-Touch, covering 464,431 interaction clips and about 1.33 billion frames across piezoresistive, magnetic and vision-based tactile sensors; roughly two-thirds came from human operators wearing tactile gloves and one-third from a dual-arm teleoperation system equipped with the Sharpa Wave dexterous hand. The model uses multi-branch encoders with a shared backbone: each of the three sensor types has its own lightweight tokenizer, and the resulting tactile tokens enter the same Haptic Transformer. The team built a unified fingertip geometric coordinate system so each tactile token carries a five-dimensional address made of a time-window index, finger number and 3D coordinates, to handle differences in sensor count, layout and sampling rate. Training follows the JEPA idea, which the team calls Tactile-JEPA: it predicts masked content in latent space rather than reconstructing raw signals, on the grounds that much of the tactile data has no contact and near-zero background signals would distract the reconstruction objective. Downstream validation is reported in three layers: five tactile understanding tasks with a frozen encoder, 26 real-robot tasks on the Sharpa Wave and XHand1 dexterous hands, and 12 simulation tasks. The report says that with only 1,000 hours of pretraining data, contact force prediction error within the 30–50 N fingertip range is below 2 N, contact center localization reaches sub-millimeter level, and contact area estimation error is below 2%. For real robots, the report says that after plugging LingBot-Touch into four action models — Diffusion Policy, π0.5, LingBot-VLA and T-Rex — as an additional tactile representation module and doing downstream post-training, average success rates generally rose; for example, π0.5 on the Sharpa Wave went from 47.56% to 68.58%, and LingBot-VLA from 45.60% to 72.51%; on XHand1, π0.5 went from 49.71% to 74.17%. The report also gives comparison data with FTP-1 and a scaling experiment increasing pretraining data from 1,000 to 5,000 hours. It lists release channels including a project page, technical report, GitHub, Hugging Face and ModelScope, but never mentions license terms, an API or pricing.

What it means for developers

The significance for developers is nothing like another large model getting an API; it needs to be viewed on three levels. First, whether you still need to build your own tactile encoder. The report's central claim is that tactile representations can be reused across sensors and across action models: the same pretrained features improved average success rates when plugged into four policy models, without retraining a tactile encoder for each action model. If that claim holds on your own hardware, what it saves is not one training run but the adaptation work of redoing things every time you switch dexterous hands or sensors. The report says LingBot-VLA had already demonstrated "one brain, many robots," and LingBot-Touch pushes the same idea into the harder-to-unify tactile modality. Second, the data cost structure. The reported source split (about two-thirds human-glove, one-third teleoperation) points to a collection path that does not depend on robot bodies: daily human manipulation can accumulate contact experience, with teleoperation filling in real interaction on specific hardware. For teams short on real-robot time, that ratio is itself a collection strategy worth discussing, not just a training detail. Third, the evaluation framing. "Adding touch helps" and "the quality of the tactile representation determines how much it helps" are two different conclusions: in simulation, the report says policies using only binary contact signals or no tactile input scored 0% on a sphere-balancing task, while LingBot-Touch scored 67.7% and 35.9% on the Sharpa Wave and XHand1 respectively. That means simply adding a contact switch will not bring equal gains; how touch is integrated, not whether touch exists, drives the result. Note that all these numbers are self-reported by the team, and the report mentions no third-party reproduction or independent evaluation.

Integration and evaluation

The two or more decision rules below are recommendations and conditional analysis based on the reported content, not factual conclusions of the report itself. Decision one (recommendation): reuse LingBot-Touch as a shared tactile encoder, or keep training a dedicated encoder for each action model. The corresponding reported change is that the same pretrained representation can plug into four architectures — Diffusion Policy, π0.5, LingBot-VLA and T-Rex — and the report says real-robot average success rates generally rose. The preconditions are strict: your tactile sensor must fall into the three reported categories of piezoresistive, magnetic or vision-based tactile — the report explicitly says pretraining does not yet cover capacitive sensors, and real-robot validation covers only the Sharpa Wave and XHand1 hands, so teams whose finger structure or sensor design differs substantially cannot directly adopt this reuse conclusion. Measurable criterion: on your own real-robot task set, compare a frozen encoder against your existing tactile baseline (e.g., a self-trained CNN tactile feature) on the same tasks, recording the success-rate difference and contact force prediction error; the reported reference magnitude is average contact force error on the same task and platform dropping from 2.31 N with FTP-1 to 0.81 N, but what you need are numbers you collect yourself, not a restatement of these. Also confirm one thing before starting: whether your sensor's measurement-point layout can map to the unified Cartesian coordinate system and five-dimensional token address described in the report; if it cannot map, all the reuse that follows is moot. Decision two (recommendation, for teams training or continuing to pretrain themselves): whether to put the next round of resources into data scale or model capacity. The corresponding reported change is the scaling experiment — as pretraining data rose from 1,000 to 4,000 hours, HaT-Tiny improved steadily across four dimensions, and at 5,000 hours it was roughly on par with 4,000 hours, leading the team to judge that model capacity was becoming the limit; switching to the larger HaT-Small brought further gains at every data scale. Precondition: you already have a stable, scalable tactile data pipeline, otherwise solve collection before backbone size. Measurable criterion: plot a "pretraining hours vs. contact force error/success rate" curve on a fixed evaluation set and see whether a plateau has appeared; if it has and more data brings no gain, then consider enlarging the backbone and rerunning the same curve for comparison. Decision three (recommendation, for data collection planning): prioritize human-glove collection or teleoperation collection. The corresponding reported change is the data source structure of roughly two-thirds glove and one-third teleoperation. Precondition: whether your target hardware is the same model as in the teleoperation system — if not, the adaptation value of the teleoperation portion to a specific body drops. Measurable criterion: after transferring the same representation to a second dexterous hand, observe the drop in average success rate; a small drop suggests a high share of transferable representation, making glove data more valuable at the margin.

Limits and open questions

First, in this catalog check (November 12, 2026), LingBot-Touch does not appear in KimiSeek's model catalog, and the report mentions no API or pricing; the project page, code repository, Hugging Face and ModelScope addresses given in the report are all published by the vendor, so any statement about whether the model is available or how to access it should defer to the vendor's (Ant Group Lingbo/Robbyant) channels, not this site. Second, the report never mentions license terms; "weight pages are publicly visible" and "commercially usable" are two different things, and you must check the license yourself before deployment. Third, real-robot validation covers only the Sharpa Wave and XHand1 hands, pretraining data does not cover capacitive sensors, and the report itself acknowledges that more robot forms, complex environments and long-horizon tasks need broader experiments. Fourth, all success rates, errors and comparison numbers are self-reported by the team, and the report mentions no third-party reproduction; for the FTP-1 comparison baseline, implementation details and tuning budget are not explained in the report either, so the fairness of the cross-comparison cannot be confirmed from the report. Fifth, teleoperation data was collected only on the dual-arm system with the Sharpa Wave, and the report gives no quantification of how large the data distribution gap is across bodies. Sixth, model parameter count, compute for training and inference, and inference latency are not disclosed in the report, so deployment cost cannot currently be estimated from public information. None of the above conflicts with the reported facts and recommendations in the body: performance numbers in the body are always labeled as reported figures, and this section describes what remains missing beyond those numbers.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

Fable 5.2 吃掉机器人?「手感」这块,还得靠具身原生基模来补全

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.