Dyna-2 Learns to Untwist Bottle Caps Out-of-the-Box from Just 13 Mins of Data

JasonMa2020 · x · 2026-08-12

The Dyna-2 model demonstrated astonishing few-shot generalization capabilities. The research team equipped the robot with a dexterous hand resembling a human's and collected only 13 minutes of real-world manipulation data for post-training.

Thanks to the extensive pre-training on 1 million hours of human video, this fine-tuned model was able to perform the complex task of untwisting real water bottle caps completely out-of-the-box. Typically, training a dexterous hand robot requires several hours of data, but Dyna-2, leveraging priors established through human video pre-training, drastically broke the dependency on robot-specific data.

Original post →

More from Embodied

• Fusion startup FAETON-X hits record 1.27e12 neutrons per shot, a first for commercial companies — dunkhippo33 · 2026-08-12
• Robot Founder Uses His Own Humanoid Robot as Post-Surgery Caregiver — CyberRobooo · 2026-08-12
• Open-Sourced Modular Jaw Upgrade for NormaCore Robot Gripper — philfung · 2026-08-12
• Horizon Robotics' Sparse4D Perception Framework Accepted by TPAMI — rsasaki0109 · 2026-08-12
• AI Makes Intelligence Cheap, the Greater Bay Area Is the Best Place to Turn Bits Into Atoms — CyberRobooo · 2026-08-12
• FEAGINE Launches Soft Robot Matrix and Cross-Embodiment Foundation Model Fi0 — 量子位 · 2026-08-12

Embodied channel →