5.5M-param tactile encoder hits real hands, lifts pi0.5 scores from 48 to ~70
Alternative_Set4042 · reddit · 2026-11-10
Last month two preprints fed GPT-6 Astra tactile images for manipulation, but both were simulation-only — one with a median 14s per model call. Robbyant now released LingBot-Touch, a 5.46M-parameter tactile encoder trained on magnetic, piezoresistive, and camera-based sensors, tested on real five-finger hands.
- Code and weights open-sourced on GitHub and Hugging Face
- Plugged into pi0.5, it bumped 26-task scores from 48 to 70 (partial credit, not binary success)
- Limits: only two hand types tested, no real palm coverage
The poster argues the two are complementary: big models plan, the small encoder handles contact.
More from Embodied
- As sims top out, the bottleneck for robot RL shifts to real-world environments — paigeinsf · 2026-11-10
- Dyna Robotics' Taku does zero-shot long-horizon whole-body manipulation live at CoRL — JasonMa2020 · 2026-11-10
- GChongkai brings DreamCan, ManiLadder & SteerAct plus a TAG/G2A oral to CoRL 2026 — chris_j_paxton · 2026-11-10
- Stanford's PsiBorg: one 8B next-token-prediction model acts as policy, simulator, tracker — chris_j_paxton · 2026-11-10
- UBTECH and China Telecom Target 100,000 Humanoid Robots in Three Years — xmercury_one · 2026-11-10
- Shield AI's Hivemind flies multiple Air Force CCAs as a coordinated team for the first time — BrettKrieger12 · 2026-11-10