JuNao PanShi launches Cog-WM 1.0 as a brain-inspired cognitive world model for embodied robots
The AMW Read
Early-stage CN robotics entrant ships a named brain-inspired world-model stack with disclosed Nav/Manip benchmarks, updating the embodied player map without yet proving segment-wide adoption.
JuNao PanShi launches Cog-WM 1.0 as a brain-inspired cognitive world model for embodied robots
Shanghai-based JuNao PanShi (具脑磐石) on September 14, 2026 released Cog-WM 1.0 at the Pujiang Innovation Forum brain-inspired embodied intelligence sub-forum, billing it as the first cognitive world model built from a structured set of brain-inspired neural mechanisms. The system pairs latent-space prediction inspired by cognitive-map memory with JEPA-style joint embedding prediction, spanning Cog-WM Nav 1.0 for map-free navigation and Cog-WM Manip 1.0 for value-guided manipulation. On disclosed figures, navigation success on an HM3D-ObjectNav subset rose from 78.50% to 86.89% versus the BSC-Nav baseline, while manipulation reportedly topped massively pretrained baselines including π0.5 by more than 16% across three major suites under a unified reproduction protocol. The stack has been shown on wheeled humanoids and quadrupeds for map-free planning, spatiotemporal memory retrieval, and spatial Q&A, with manipulation verified on a wheeled humanoid and Nav 1.0 already used in quadruped patrol customer sites.
The launch matters because it argues that embodied progress is blocked less by engineering polish than by paradigm: stacking data and parameters does not automatically buy physical-world cognition. By stressing goal-conditioned latent prediction, structure-content memory separation with surprise updates, and multi-horizon foresight, JuNao PanShi is pitching algorithm mechanism as a substitute for ever-larger scene corpora—an early stake in the fight over whether embodied systems need foundation-scale pretraining or structured world models. Founded in June 2025 and tracked in the AI Market Watch index at about $20.0M total funding among roughly 5,000 covered companies, the firm is still proving that path outside lab protocols.
For builders and investors, the near-term test is third-party replication: whether multi-timescale prediction and value-modulated experience learning keep their edge against π0.5-class policies and map-based SOTA when lighting, viewpoint, and background shift in open facilities. Watch whether published arXiv methods and ablation ladders (for example, 80.0% to 84.6% on LIBERO-Plus as mechanisms stack) convert into durable deployment economics, and whether the company's LeCun AMI Lab-aligned positioning draws partners across brain-inspired algorithms and neuromorphic silicon in China's embodied stack.
