Dyna-2

Headlines · 3

1. 3d
   Nature Will releases Natura-1: 50B MoE model for dexterous robot hands

   Beijing embodied-AI company Nature Will released Natura-1, a 50B-parameter MoE Transformer physical-world foundation model pretrained on roughly 1 million hours of first-person human experience data (about 61.7 million clips, 318,000 tasks, 14,000 scenes), then post-trained on limited robot data; it takes language and vision as input and drives 20-degree-of-freedom dexterous hands within a 100 ms control loop. The team says it combines dexterous manipulation with zero-shot instruction following, and published unedited one-take demos including wrapping wontons, making tea, packing parcels, applying a phone screen protector and writing calligraphy. On its own 12-dimension benchmark, Natura-1 scored 53.4 (Base) and 59.1 (Agentic) on the dexterous-hand platform, while π0.5 and OpenAI's GPT-6 Astra scored zero there (13.5 and 16.7 on a gripper platform); the team also reports action-prediction error dropping 47.7% on held-out human data and 67.7% on held-out robot data as training experience scaled from about 44,000 to 1 million hours.

   机器之心 · 🔥 0
2. 3d
   VPP2 and Dyna-2: how human video trains robots

   A TMTPost feature by Jushenzhi reports that on October 10 Robot Era open-sourced two sets of VPP2 video model weights; VPP2 splits manipulation learning into predicting how a task unfolds, compressing it into visual plans and learning robot actions, and only the first two stages were released. Dyna Robotics' Dyna-2 is described as a scale test: four models pretrained on 1,000, 10,000, 100,000 and 1 million hours of first-person human video then received identical robot-task training, the first three failed to turn a key while the 1-million-hour model succeeded 9 of 10 times. Nvidia's DreamZero added 12 minutes of action-label-free human video for nine previously unseen tasks and lifted average task progress from 38.3% to 54.3%, while RoboWM-Bench from Peking University and others converted generated video into robot actions in simulation, reporting 50%, 20% and 0% success for Wan 2.6 on closing drawers, picking objects and placing objects. The article argues human video is taking over task understanding and world dynamics while scarce robot data focuses on body control.

   钛媒体 · 🔥 0
3. 3d
   Dyna-2.1 physical robot agent folds napkins 24 hours straight at 99.4% success

   In a Latent Space podcast interview, Dyna Robotics co-founder and chief scientist Jason Ma described Dyna-2, the world-action model the company announced last month, and the physical robot agent that followed, Dyna-2.1. Dyna-2 was pre-trained on more than one million hours of human video and predicts both future video and robot actions, with the company claiming new scaling laws on human data across four orders of magnitude. According to the interview, Dyna-2.1 reached a 99.4% success rate folding restaurant napkins over 24 continuous hours—recovering even after pulling over the whole stack—and is said to fold thousands of napkins daily at Din Tai Fung, while robots can learn new tasks from as little as 45 minutes of demonstrations.

   Latent Space · 🔥 0

Experience and discussion from the community

Share my Dyna-2 experienceAsk about Dyna-2