robot learning

World Models

2026-09-11 02:56:12

What counts as a world model? Fei-Fei Li, LeCun and Zhu Jun offer three different answers

The term "world model" has become one of the least settled concepts in AI in 2026. It can refer to a model that generates coherent video, a digital environment that keeps responding to user input, a latent-space system that predicts future states, or a policy that outputs robot actions. Those systems all process information about the world, yet they do not solve the same problem, and comparing them by visual quality, geometric consistency, prediction accuracy or task success can blur more than it clarifies. A MarsBit article, citing work from World Labs, Meta and Tsinghua University professor Zhu Jun’s team, lays out three leading interpretations. Fei-Fei Li and World Labs sort world models into renderers, simulators and planners. Yann LeCun argues for learning predictable structure in an abstract latent space. Zhu Jun and his collaborators define a broader "general world model" from first principles, built around three linked capabilities: understanding, imagination and action. Their framework also proposes a five-level roadmap from L1 world generation to L5 world organization, and places video generation, real-time interaction and embodied control on a single capability curve. The piece further examines a data pyramid spanning internet video to robot interaction, the MoT architecture for multimodal coordination, and Motus2, which combines policy generation, simulation, evaluation and tactile feedback in one closed-loop system.

840

Agibot

2026-08-03 11:45:08

Agibot removes chief scientist Luo Jianlan from partner roster as departure speculation grows

Agibot has removed Luo Jianlan from the partner team list on its official website, adding weight to online speculation that the company’s former chief scientist may have left. Luo had previously been listed as partner, senior vice president and chief scientist, and appeared on the company’s first publicly disclosed partner roster in September 2025 alongside Deng Taihua, Peng Zhihui and Yao Maoqing. Other public-facing changes have surfaced as well: Luo’s personal webpage now only identifies him as an assistant professor at the Shanghai Institute of Advanced Intelligence, while his X profile no longer includes Agibot-related titles. The timing has drawn attention because Zhiyuan Innovation said on July 24, 2026 that it had started the process for a Hong Kong listing. Agibot and Luo had not publicly announced any personnel change as of publication. Luo joined Agibot in April 2025 as chief scientist, led the creation of the company’s Embodied Intelligence Research Center, and later became a partner and senior vice president. His past work spans robot learning, real-world reinforcement learning and embodied intelligence, with prior research roles at UC Berkeley, Google X, BAIR, and projects including SERL and HIL-SERL.

450