On October 28, Zhiyuan released an upgraded version of its embodied intelligence simulation platform Genie Sim, focusing on improving soft-body simulation and adding model evaluation and reinforcement learning post-training features. In comparative tests covering 4 robot platforms and 22 tasks, the correlation between simulation and real-robot task scores reached 96.7%. Related code is planned to be updated on GitHub, with simulation assets and datasets to be opened to developers.
On October 27, 2026, RoboDojo, the embodied intelligence evaluation benchmark led by the University of Hong Kong, released its latest ranking. Zhiyuan GO-3 Nano Preview ranked first with an average score of 61.90 and an average success rate of 56.42%, surpassing GPT-6 Astra, and also topped the sub-dimensions of fine manipulation, long-horizon execution and open semantic understanding.
Chinese robotics company AgiBot made 110 full-size Expedition A3 humanoid robots dance in sync in Jinan, with a performance lasting nearly six minutes featuring jumps, flips, wave patterns, and formation changes, reportedly earning a Guinness World Records title. All robots were production units, coordinated to the centimeter via the LinkBots control system and ultra-wideband positioning.
On October 27, Zhiyuan officially launched Genie Master, an embodied agent framework. The Spirit G2 robot equipped with the framework has been working at a real convenience store without manual demonstrations or debugging, achieving continuous self-evolution with no human in the loop, and marking the first validation of Physical RSI (embodied recursive self-evolution) in a real business scenario.
Zhiyuan today released its safety-certified heavy-duty embodied robot Lingjing G2 Max for industrial scenarios such as depalletizing, material transfer, and production line feeding. The robot has a 40kg rated dual-arm payload and has been running on multiple production lines, achieving over 99% success rate in e-commerce warehousing tasks.
Jensen Huang announced at the GTC Berlin event that NVIDIA is partnering with Agile Robot to release a robot, planned for the first quarter of 2027. The article uses this as a starting point to analyze the layouts of frontier AI companies such as OpenAI, Anthropic, and Google DeepMind in embodied intelligence, and the impact of GPT-6 Astra on the industry.
On October 22, Zhiyuan officially open-sourced the fourth phase of the AGIBOT WORLD 2026 dataset, focused on embodied reasoning. The dataset constructs decision-making scenarios for the same task under different states based on object quantity, occlusion, position occupancy, and task progress, covering nine core reasoning abilities including attribute selection, condition judgment, action feasibility evaluation, target search, long-horizon planning, and process memory, with Easy, Medium, and Hard difficulty levels.
A joint team from Tsinghua University, AgiBot and Infinigence AI proposed Collective Experience Learning (CEL), which lets deployed robots share execution experience and improve policies together. In five real-world tasks, using 96 robots improved phone packaging success from 79.2% to 98.3%.
In an interview on October 22, 2026, Google DeepMind CEO Demis Hassabis compressed his AGI timeline from three to five years to two to three years, citing accelerating progress. He reaffirmed AGI's primary role in accelerating scientific discovery and maintained a cautious regulatory stance, while noting Isomorphic Labs' $2.1 billion Series B and plans to file its first clinical trial application by end of 2026.
At the GTC Berlin conference, NVIDIA expanded its Isaac robotic development platform and, together with Agile Robots, released the humanoid robot reference platform Agile ONE SN, scheduled for early 2027, featuring the Jetson Thor platform with 2070 TFLOPS of compute. The company also launched the open-source frameworks Isaac Lab 3.0 and Isaac OS, along with three reference workflows and resources such as the SimReady Central Library and Human 1M dataset.