Ant Group's Robbyant open-sources LingBot-Touch, a tactile model trained on nearly 5,000 hours of data

Robbyant, the embodied-AI company under Ant Group, has open-sourced LingBot-Touch, a tactile foundation model meant to give robot hands one shared sense of touch. Announced on November 10, it was pretrained on nearly 5,000 hours of real tactile data and covers three sensing principles: magnetic, visuotactile and piezoresistive. The code is released under Apache-2.0, with Tiny and Small checkpoints on Hugging Face and ModelScope.

The problem: incompatible sensors

Tactile sensors produce incompatible raw signals, so touch experience gathered on one sensor rarely transfers to another. LingBot-Touch gives each sensor type its own input tokenizer feeding a shared Transformer. It maps contact points into a common fingertip coordinate space and outputs unified tactile tokens that different action models can use.

The repository describes a JEPA-style objective that predicts latent representations from spatiotemporal touch context. Training data came from both human-worn tactile gloves and teleoperated dexterous hands.

Results (Robbyant's own)

Robbyant tested the model on two dexterous hands, Sharpa Wave and XHand, across 26 real-robot tasks including page turning, bottle-cap twisting and identifying mahjong tiles. With LingBot-Touch's representation added, four action models improved average task success by about 14 to 26 percentage points over versions without touch. The four models were LingBot-VLA, π0.5, Diffusion Policy and T-Rex.

In one case, T-Rex twisting caps on Sharpa Wave succeeded 56.25% of the time without touch, 62.5% with raw sensor signals and 90% with LingBot-Touch.

Model size and cross-hardware transfer

All 26 tests used the Tiny version, at roughly 5 million parameters. It adds about 0.13% to 0.16% to the policies it plugs into. The team says the larger Small version kept improving as pretraining data grew toward 5,000 hours. Parameter count for Small is not stated.

To test cross-hardware transfer, Robbyant kept the encoder frozen on XHand, whose magnetic sensor model was not in the pretraining data, and trained only the action model on XHand demonstrations. Its magnetic-touch knowledge there came solely from human glove data.

Release details

• License: Apache-2.0
• Checkpoints: Tiny and Small, on Hugging Face and ModelScope
• Scope of GitHub release: inference only, containing the frozen encoder and utilities for extracting tactile features
• Pricing: not applicable to the open-source release; no hosted pricing disclosed

All results are Robbyant's own, and the company has not reported independent replication.

What this means

Robbyant positions LingBot-Touch as a vendor-neutral layer between tactile sensors and action models, not a module tied to its own LingBot-VLA. Four action models from different developers were tested with it, which fits that framing. Per Robbyant, the intended benefit is that hardware makers can pick sensors by cost while system teams keep their own policies.

The XHand experiment is the most relevant test of the shared-representation claim, since that hand's magnetic sensor was absent from pretraining. Because the released code covers only the frozen encoder and feature extraction, users can plug the encoder into existing policies but cannot reproduce pretraining from this release. Independent replication will determine whether the reported gains hold beyond Robbyant's two test hands and 26 tasks.