Knowin's GLOW Teaches Robots From One Demonstration, Tops Three Embodied AI Benchmarks
新智元 · wechat · 2026-09-24
Startup Knowin unveiled GLOW (Generative Learning of World), a natively unified autoregressive model where vision, language and action share one sequence—BrainSkill for understanding, ActSkill for closed-loop control.
Highlights
- One-shot learning: robots generalize box-packing, watering, wiping, cocktail-pouring from a single human demo, without weight updates
- Harness feedback system gives every action an execution receipt; GLOW opened a stuck drawer in 1138 steps vs GPT-6 Astra's failed 1200, and completed a bottle-to-bowl task in 157 steps vs 863
- KnowinDream generates 200M synthetic episodes; real execution results flow back to calibrate the world model, forming recursive self-improvement (RSI) with safety regression gates
- Benchmarks: 62.2% on RoboDojo (vs 22.2% for GPT-6 Robocurve), 86.7% on LIBERO-Pro, #1 on EmbodiedArena 2D
- The X1 home robot (under 40cm folded) launches next year
More from Embodied
- Robots fold laundry and write code, but camera calibration still takes hours — RexDouglass · 2026-09-24
- Liquid AI partners with Qualcomm to bring on-device 'Liquid Context' to Snapdragon — JosephJacks_ · 2026-09-24
- 36B Agent Model Runs Fully Local on Snapdragon X2 Elite With Just 32GB RAM — Kyrannio · 2026-09-24
- EgoSuite and continuous learning for robotics shown at Humanoids Summit Seoul, next at IROS — jonstephens85 · 2026-09-24
- CUHK's PackLab: an MLLM framework for closed-loop robotic bin packing beats RL — CUHK-CSE · 2026-09-24
- Alexandr Wang Shows Two Robots Playing Tennis — alexandr_wang · 2026-09-24