LingBot-Touch

LingBot-Touch is an open tactile foundation model from Robbyant (Ant Group) for dexterous robot manipulation. Announced around 9 November 2026, it uses a JEPA-style objective with a shared Haptic Transformer pretrained on roughly 4,983 hours and 1.33 billion frames spanning piezoresistive, magnetic, and visuotactile sensors.

Tiny and Small weight variants are published on Hugging Face. The model is designed to transfer across sensor types and dexterous hands, so labs can reuse one tactile backbone instead of training from scratch per hardware stack. Product details live at https://technology.robbyant.com/lingbot-touch.

Highlights:

• Official Ant / Robbyant open-source launch with product page + HF weights;
• Haptic Transformer / JEPA pretraining on multi-sensor tactile corpora (~5000h);
• Cross-sensor and cross-hand transfer for real-world dexterous tasks;
• Tiny & Small checkpoints for different compute budgets;
• Aimed at boosting contact-rich manipulation beyond vision-only policies.

Why it matters

Vision-only robot policies still struggle when contact, slip, and force matter. A tactile foundation model pretrained at multi-thousand-hour scale gives teams a shared representation for piezoresistive skins, magnetic tactile arrays, and visuotactile cameras—reducing the need to collect and train a new encoder for every hand.

Robbyant’s open release (docs + HF Tiny/Small) makes the work reproducible for research and industrial pilots. Cross-sensor transfer is the practical hook: if one backbone adapts across modalities, fleets with heterogeneous hands become easier to support. That fits the broader push toward foundation models for robotics (VLA, world models, and now haptic FMs).

Limits

Transfer quality still depends on calibration, mounting, and how close your sensors are to the pretraining distribution. Tiny/Small are research/prod-preview weights—not a guarantee of every dexterous benchmark. JEPA-style pretraining helps representation learning but downstream policies still need task data and safety constraints for real robots. Confirm license, HF model cards, and Ant/Robbyant terms before commercial deployment. Do not confuse LingBot-Touch with unrelated “LingBot” chat or language products.

1. Read https://technology.robbyant.com/lingbot-touch and the Hugging Face model cards for Tiny/Small;
2. Match your tactile sensor type (piezoresistive / magnetic / visuotactile) to the supported adapters;
3. Load the backbone, fine-tune or probe on your hand’s contact tasks, and evaluate slip/force-sensitive skills;
4. Before production, verify license, robot safety interlocks, and that claimed cross-sensor transfer holds on your hardware.