AI BriefingRobbyant's LingBot-Touch Outperforms Reconstruction Models in Tactile Pretraining
·2026.11.10 03:52
Key point
The self-supervised encoder achieves 49.3% tactile-to-video recall at 4K hours, significantly outperforming reconstruction-based baselines.
Details
- Robbyant releases LingBot-Touch, a self-supervised tactile encoder trained with JEPA-style latent prediction rather than raw signal reconstruction.
- At 4K hours of data, the JEPA model reaches 49.3% top-5 recall, while MAE stays under 9% and VAE drops to 3.1%.
- The approach addresses the issue where reconstruction loss shrinks on near-zero background signals without learning actual contact features.
- Frozen encoder features scored highest in 89 of 104 policy and task combinations across 26 real robot tasks.
- A magnetic hand not seen during pretraining saw performance improve from 57.21 to 75.04 using the LingBot-VLA model.
- Training utilizes a masking ratio warming up from 40% to 80% over 300 steps to prevent gradient explosions in sparse signals.
- Three sensor types (magnetic, piezoresistive, visuotactile) feed a shared Transformer using 5D RoPE for time, finger, and 3D position.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI Disbands Robotics Team, Pivots to General Models for Embodied AI
Runway Releases Praxis-1 Open-Weight Robotics ModelNVIDIA Releases GR00T 2 Foundation World-Action Models for Cross-Embodiment Robotics