XNews

Robbyant Releases LingBot-Touch Open-Source Tactile Foundation Model

Robbyant Releases LingBot-Touch Open-Source Tactile Foundation Model

Robbyant released LingBot-Touch, an open-source tactile foundation model pretrained on nearly 5,000 hours of real tactile data. The JEPA-based model uses a shared Haptic Transformer to unify magnetic, piezoresistive and visuotactile signals, and it improved average task success by roughly 14 to 26 percentage points across four action models on 26 real-robot tasks.

Event version 1.

Key insights

Unified heterogeneous tactile architecture

LingBot-Touch jointly models magnetic, piezoresistive and visuotactile signals using JEPA-based latent prediction and a shared Haptic Transformer. It encodes inputs as unified tactile tokens, enabling multi-sensor integration and reusable features for downstream action models.

Real-robot gains across four action models

Evaluation covered 26 real-robot tasks spanning force control, deformation, visual occlusion and bimanual coordination, using Sharpa Wave and XHand dexterous hands. Average task success gains ranged from approximately 14 to 26 percentage points over the no-touch baseline across LingBot-VLA, π0.5, Diffusion Policy and T-Rex.

Simulation results and scaling

Across all 12 simulation tasks on Sharpa Wave and XHand, LingBot-Touch achieved the highest success rates and cumulative returns. As pretraining data grew to nearly 5,000 hours, the Small variant continued to improve in tactile perception tests.

Two open-source model sizes

Open-source Tiny and Small variants balance representational capacity with lightweight deployment for high-frequency tactile perception. The HaT-Tiny encoder has 5.46 million parameters.

Pretraining corpus composition

The pretraining corpus comprises 4,983 hours of recordings, 464,431 interaction episodes and roughly 1.33 billion frames. It includes 2,197 hours of piezoresistive, 1,136 hours of magnetic and 1,650 hours of visuotactile data.

Source posts