2077AI

2077AI papers at #CoRL2026 on Thursday, Nov 12: TouchScale and ENCORE, presented across three workshops. 🧤 TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning • Scaling H2R Workshop, poster session 10:30–11:00 AM CST • Everything Beneath the Policy, poster session 10:35–11:05 AM CST, Hilton Austin, Salon E 500 hours of human vision and touch data from one wearable setup. TouchScale mid-training improved real-world robot success from 22.5% to 57.5% across four tasks. 🔗 touch-scale.github.io 💻 ENCORE: Few-Shot Agentic Discovery of Manipulation Strategies • Workshop on Agentic Robotics, poster session 4:20–5:55 PM CST A coding agent reads a few demonstrations, writes a policy program, and freezes it. 96.3% on LIBERO-PRO perturbation tasks vs. 89.3% for the strongest prior agentic system (same model). 🔗 yifank.github.io/encore Come check out our works at the poster sessions!

What tells a simulator how hard a human grips? Video shows where each finger stops. How far it closes, which sets grip force, goes unseen, so prior pipelines that rebuild demos from vision never get contact force. TouchTwin, a CoRL 2026 late-breaking result built on TouchScale, adds it: selected recordings replayed in MuJoCo, with the glove's measured touch setting each finger's force. On held-out recordings: • A hand policy's grip force, in sim: 0.31× the recording when trained on observed pose, 0.99× on touch-grounded replays • A contact predictor misses 13% of real contact frames when trained on touch-grounded replays, vs 32% on vision-based replays Limits: policy tested in simulation only; mug task only; small test sets. 📍 Presented today at CoRL: Real2Sim2Real poster session, lunch break, Griffin (2nd floor), JW Marriott Austin. A collaboration between Stanford, Texas A&M, Yale and 2077AI.

Proactive agents face a calibration problem: acting correctly is not enough if the intervention itself is poorly timed or unwarranted. We introduce HALO, a human-grounded benchmark for evaluating proactive agents across everyday environments. 📄 arxiv.org/abs/2610.39086 🧵 (1/5)

ENCORE (Oral, NeurIPS 2026 Robot Learning Workshop; CoRL 2026 Agentic Robotics Workshop): a coding agent reads a few robot demos, writes a policy program, refines it, then freezes it. 96.3% on LIBERO-PRO, vs 89.3% for the best prior agent on the same model. arxiv.org/abs/2609.37359 yifank.github.io/encore

⚔️ We are pleased to introduce CyberArena, where we benchmark 12 frontier AI agents via letting them combat with each other in live attack-and-defense security contests. 🧵