Gen‑0 vision model enables UR5 packing
A generalist vision foundation model called Gen 0 (trained on ~250k hours of data) has been used to enable UR5 arms to autonomously pack phones after human fine‑tuning, showing foundation models moving into low‑level manipulation tasks. This is an example of large vision models being repurposed for embodied robotics workflows. (x.com)
Generalist publicly showcased GEN‑0 performing an autonomous smartphone‑packing demo at NVIDIA GTC, calling it their “first live public demo” on the company site. (generalistai.com) GEN‑0 was announced in a Nov. 4, 2025 research post and Generalist says it was pretrained on an in‑house dataset of over 270,000 hours of real‑world manipulation data that the team reports is growing by ~10,000 hours per week. (generalistai.com/blog/nov-04-2025-GEN-0) The team describes a “Harmonic Reasoning” objective that interleaves sensing and acting, reports an empirical phase transition at ~7 billion parameters where smaller models “ossify,” and says they have scaled GEN‑0 past 10 billion parameters. (generalistai.com/blog/nov-04-2025-GEN-0) At NVIDIA GTC 2026 Universal Robots and Scale AI demonstrated the same smartphone‑packing task on production cobots, with two UR7e robots autonomously executing the packaging workflow during the event. (robotics247.com/article/nvidia-gtc-2026-universal-robots-scale-ailaunch-imitation-learning-system-to-accelerate-ai-model-training) Universal Robots’ UR AI Trainer used a leader‑follower imitation setup that captures synchronized motion trajectories, force feedback and visual data on the target cobots, and the company highlighted Direct Torque Control plus force feedback as the fidelity advantage for contact‑rich tasks. (thenextweb.com/news/universal-robots-and-scale-ai-launch-the-ur-ai-trainer) Generalist’s blog notes that GEN‑0 shows “fast adaptation to new tasks with increasingly less post‑training,” and the GTC demos position embodied foundation models as a pathway to transfer large‑vision/model capabilities onto industrial cobots that already power over 100,000 deployments. (generalistai.com/blog/nov-04-2025-GEN-0)
