Ant's Robbyant trains LingBot-Touch: 2.5h of labels teach robot hands to sense force
heyshrutimishra · x · 2026-11-10
Ant Group's embodied AI unit Robbyant released LingBot-Touch, a tactile foundation model pretrained on 4,983 hours of real touch recordings using JEPA-style masked latent prediction — no frame-level force/contact labels.
- After freezing the 21.76M-param encoder, a probe trained on 2.5 hours of force labels reads fingertip force with 0.81 N error on the Sharpa Wave hand (0.12 mm contact-center error), and 1.32 N on the unseen XHand1 sensor.
- The deployed 5.46M-param Tiny version lifts mean success on 26 real robot tasks by 22–25 points when plugged into π0.5, T-Rex, and LingBot-VLA.
- Caveat: all numbers come from Robbyant's own report (20 trials/task, two hands); third-party validation pending. Touch can't be scraped off the internet, making it a scarce modality.
More from Embodied
- Anthropic touts Fable 5.2 robot control gains on RoboDojo; real-world box folding disappoints — _Stocko_ · 2026-11-11
- AI2 open-sources ToyBlocks to benchmark robotic world models by task-relevant fidelity — allen_ai · 2026-11-11
- OpenTac: open-source ecosystem turns any robot surface into a tactile skin — _krishna_murthy · 2026-11-11
- NVIDIA launches open Alpamayo 2 Edge for real-time in-vehicle AI reasoning — nvidia · 2026-11-11
- NVIDIA open-sources GR00T 2 robot foundation model with 100K hours of training data — DrJimFan · 2026-11-11
- Stanford DroneHacks: 100 students fly GPS-denied PX4 drones on stereo cameras in 48 hours — mrpollo · 2026-11-10