Coowa breaks down COOWAM: on-device world model behind 10,000+ deployed city robots
创业邦 · wechat · 2026-09-21
- Coowa, with 10,000+ robots across 50+ cities (RaaS orders >¥5B, ¥1B revenue, EBITDA positive, >$600M raised), details its COOWAM dual-layer world model: UrbanVLM for slow deliberative planning, CooWAIM for fast execution, built on a System1/System2 dual-core.
- The model runs natively on two NVIDIA GPUs per robot, forecasting only 0-4 seconds ahead. Latent-space inference cuts compute 75% (sub-100ms latency); a rollout/stop gating allocates compute by risk-benefit; AIGC-generated counterfactual danger videos lifted risk-detection AUC from 0.5 to >0.93.
- The piece also surveys six world-model schools (LeCun's JEPA, Fei-Fei Li's spatial intelligence) and the field's data-flywheel and commercialization challenges.
More from Embodied
- ZuckOff App Detects Nearby Meta Smart Glasses via Bluetooth Fingerprints, Tops 5,000 Downloads — LexiLove · 2026-09-21
- GaME (CVPR 2026): Gaussian mapping that forgets stale geometry as robot scenes change — lucacarlone1 · 2026-09-21
- Stanford RL method fixes VLA latency, lifting robot success from 42% to 97% with 10 minutes of data — burny_tech · 2026-09-21
- Deadlifting 160kg at LUDUS: a striking strength demo — AIFlow_ML · 2026-09-21
- DeformSmith generates physics-grounded deformable assets for robot manipulation — Can Li · 2026-09-21
- Tesla's Texas Optimus line nears structural completion; Qiyuan claims a robot every 2.5 minutes — 创业邦 · 2026-09-21