Shengshu Tech Unveils 5-Stage Roadmap for General World Models
生数科技 · wechat · 2026-08-24
Zhu Jun, founder of Shengshu Tech, outlined a 5-stage evolutionary roadmap (L1-L5) for General World Models at the World Robot Conference, emphasizing a closed-loop system of understanding, prediction, and action.
Core Definition
A general world model integrates understanding, prediction, and action into a coupled feedback system, rather than acting as a standalone generator or policy model.
Technical Pillars
- Data: A multi-layer pyramid ranging from web videos to robot interaction data, utilizing "imperfect data" (failures/corrections) for learning.
- Architecture: MoT (Mixture-of-Transformers) unifies image, video, language, and robotic action processing.
- Compute: Balances large-scale pre-training with real-time inference acceleration.
5-Stage Roadmap
- L1 World Generation: Achieved via the Vidu video model.
- L2 Interactive World: ViduS1 enables real-time voice interaction.
- L3 Actionable World: Motubrain unifies environmental understanding with action generation, topping the RoboTwin2.0 benchmark and verified on nearly 10 robot types.
- L4 Autonomous World Agent: Capable of long-term planning and active exploration.
- L5 World Orchestrator: Coordinates multi-agent collaboration for complex tasks.
Future work for L4/L5 includes overcoming six challenges, such as learning physical laws, persistent memory, and online learning.
More from Embodied
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark laptops — The Verge AI · 2026-10-15
- John Deere's second-gen autonomous tillage: 16 cross-checking cameras beat sensor drift — lukas_m_ziegler · 2026-10-15
- Europe's Application-Layer Bet: Can the Rotterdam Humanoid Center Deliver? — ingliguori · 2026-10-15
- VC thesis 'Data at the Edge': cheap sensors plus AI are rebuilding the internet's data flywheel offline — rebeccakaden · 2026-10-15
- PingpongDuo shows humanoid robots playing doubles table tennis — arankomatsuzaki · 2026-10-15
- Robotics data gap: just 300K hours of manipulation data globally vs 100M+ needed, $3B spend projected — 0xSammy · 2026-10-15