BioMARS 2.0 is an embodied AI system that connects paper reading, experiment planning and robotic execution, allowing robots to help choose the next research step.
OpenAI has reportedly dissolved its robotics team, prompting debate over whether embodied AI needs a dedicated robot brain or can rely on general-purpose foundation models.
Changcheng Fund's Yu Huan says the robotics sector's recent rally is different because large-model capabilities are moving from language understanding into motion control. Overseas testing and industrial progress may signal an inflection point for embodied AI.
In a Leaderobot conversation, Lu Zongqing and Wu Zuxuan examined how world models and vision-language-action systems are evolving, and what that means for the next stage of embodied intelligence.
On October 27, an embodied AI evaluation benchmark led by the University of Hong Kong released its latest rankings, with Zhiyuan's model taking first place with an average score of 61.90 and a success rate of 56.42%, surpassing general large models. The results indicate a growing division of labor between general and embodied models, while the benchmark provides an objective capability standard for the industry.
Unitree's Wang Xingxing demonstrated GPT-6 controlling robots with zero training to draw and play Go, while outlining plans for 10 million hardware units a year and self-sufficient physical-AI compute. The market, however, still values Shoucheng Holdings (0697.HK) through its robotics investment rather than this industrial shift.
Wang Xingxing demonstrated GPT-6 controlling robots without prior training to paint and play Go, while outlining plans for 10 million hardware units a year, physical-AI model deployment, and self-sufficient compute and chips. The market still values First Cheng Holdings through a conventional robot-investment lens.
Daxiao Robotics has released ACE-Brain-0.6, a 9B embodied foundation model trained on about one million verifiable spatial-anchoring samples. It beat OpenAI's GPT-6 Astra on six of eight grounding evaluations and averaged 74.53 on seven public benchmarks versus 73.68.
Daxiao Robotics has released ACE-Brain-0.6, an embodied base model built on Qwen3.5-9B and trained with about one million spatial-anchoring samples. The company says it outperforms OpenAI's GPT-6 Astra on six of eight grounding benchmarks.
Daxiao Robotics has released ACE-Brain-0.6, an embodied foundation model fine-tuned from Qwen3.5-9B with about one million spatial-anchoring samples. The company says it outperforms OpenAI's GPT-6 Astra on six of eight grounding benchmarks.
Neura Robotics has launched Neuron AI, a perception-language-action model that integrates 3D vision, touch, force sensing, audio, and physical state into a unified robotic brain. Backed by $1.4 billion in funding and support from Amazon, NVIDIA, and Qualcomm, the company aims to supply a general-purpose AI layer across multiple robot forms via its Neuraverse platform.
NEURA Robotics explained how its Neuron AI model, a Perception-Language-Action (PLA) system, lets a robot combine several senses at once: 3D images, tactile and force signals, audio, its own position and movement, plus language, all flowing into one shared, continuously updated picture of the situation that drives its next decision. The company said the second building block is memory, noting that humans develop object permanence in their first two years of life.
Unitree founder Wang Xingxing said the release of GPT-6 has strongly influenced the embodied AI industry, and he now expects the sector's tipping point could come within one to two years.
Knowin released GLOW, a generative learning architecture for general embodied intelligence, aiming to let robots learn tasks from a single demonstration. The technical report details a unified multimodal autoregressive model and reports leading results on RoboDojo, LIBERO-Pro, and Embodied Arena.
At a Sept. 23 seminar in Beijing, ModelBest engineering vice president Zou Zhensheng said the last mile of embodied AI scaling depends on reliability and whether customers can justify the economics. He outlined data-mix strategies, a scenario ladder and ecosystem collaboration, predicting that over the next two to three years edge models will advance along a perception-thinking-action loop, with automotive, consumer electronics and robotics becoming major commercialization scenarios in turn.
Huxiu AI100 closed-door meeting gathered experts to discuss whether world models and GPT-6 can serve as robot brains. They debated evaluation criteria, language vs. action, real-time limits, self-evolution, hardware coupling, and the missing AI Coding scenario for embodied AI.
GPT-6 Astra achieved 95% success in a robot-control task with far fewer tokens than Fable 5.1, but RoboDojo tests found it issued unsafe actions and damaged hardware, forcing early termination. Noematrix's RoboRSI framework uses four-agent loops and code solidification to improve continuous execution, diagnosis, and adaptation, raising LIBERO pass rates and cutting token use.
Tsinghua University, Infinigence, and Zhengxing Innovation have jointly open-sourced RPent, an embodied agent infrastructure that connects general large models with expert VLA models, memory, tools, and robot interfaces. RPent achieves 92.6% success on LIBERO-PRO and over 7x end-to-end speedup, enabling robots to understand tasks, invoke capabilities, and adapt in open-world scenarios.
Anthropic has reportedly established a biology lab for physical experiments to advance drug discovery, aiming to have its Claude AI control robots and run lab tests with minimal human involvement, according to Reuters.
Galaxy General chairman Yin Fangming used the SCO Digital Economy Forum to argue for an open-collaboration path in embodied AI, as OpenAI spends billions to build a commanding data lead. Meanwhile a team tied to Galaxy General's CTO produced the first systematic test of GPT-6 Astra's capability limits in robotic manipulation.
On September 8, 2026, AI firms pivoted from model scale to engineering reliability and real-world adoption, with DeepSeek hiring 150 engineers (none in research), ByteDance advancing spatial video models, and Mistral raising €3B.
At its semi-annual briefing, Moore Threads confirmed supply chain readiness and announced the successful end-to-end training of a 236B MoE model and embodied AI models like RoboBrain2.5 and EvoPhys-World on its MTT S5000 cluster.
At an ICRA panel, researchers explored the overwhelming growth of robotics papers and potential responses, including LLM-based tools and peer review reform. Concerns over hallucination, salami slicing, and visibility overshadowing value were central to the debate.
OpenClaw, an open-source AI assistant, has released version 2.0, shifting its focus from standalone intelligence to cross-device agent infrastructure, with upgrades across conversation, browser control, memory, automation, plugins, and security.
StartLux's 27B-parameter local model ranked second in CAICT's MCP test, outperforming DeepSeek-V4-Flash and nearly matching the 1.6T-parameter DeepSeek-V4-Pro. The company plans to launch commercial solutions within the year.
Skild AI found that current VLA models must be post-trained with 50-100 hours of data collection and fine-tuning to match the accuracy of S1, which achieves the same with just one example of prompting.
Skild AI highlights S1's behavior: it does not simply replay video demonstrations, but uses common-sense understanding to withstand perturbations, improvise after mistakes, and sometimes execute tasks with more precision than the human demonstrator.
Skild AI says S1 is a step-change improvement over conventional VLAs: it matches language-prompted VLA performance on known tasks, and exponentially outperforms them on novel tasks as pre-training scales. The company sees this as a path toward scaling laws for robotics.
Skild AI explains that S1 learns new tasks in a manner similar to a language model: given a video demonstration as a prompt, it generates robot actions to accomplish the task in any environment and embodiment.
Skild AI recounts the first time S1 flipped a pancake. The team initially assumed the skill was in pre-training data, but a full search found no examples. S1 inferred the out-of-distribution task from a single video prompt.
Skild AI says its AI system can be taught extremely long-horizon tasks over 10 minutes long. By composing skills learned in pretraining or creating new ones, it makes coffee, pots a plant, fries pancakes, and more, covering tasks never seen during pretraining.
Shengshu Tech's founder Zhu Jun proposes a five-level roadmap for general world models, with the company covering the first three levels. His team's Motubrain model scores 96.1 on the RoboTwin 2.0 benchmark and adapts to new robot platforms with just 50-100 human demonstrations.
SpaceX AI has unveiled a Browser Use plugin for Grok Build, enabling real browser automation with local Chrome or cloud browsers. The tool supports web scraping, form filling, screenshots, and full workflow automation, with local execution via uvx requiring no API key for local Chrome.
At the 35th International Joint Conference on Artificial Intelligence in Bremen, international experts praised China's rapid advancements in AI research and applications, particularly in robotics and large language models, placing it among global leaders.
At the 2026 World Robot Conference, Moore Threads founder Zhang Jianzhong said embodied intelligence is approaching a 'ChatGPT moment,' with AI computing clusters inevitably scaling from thousands to hundreds of thousands of GPUs.
Anthropic will allow enterprise clients to keep their data on their own cloud infrastructure while using its advanced AI models, reversing an earlier policy meant to prevent cyberattacks. The new system, launching later this year, maintains a 30-day retention requirement but offers greater data control to customers.
Xiaomi announced that its MiMo-V2.5 model ranks first in weekly calls on OpenRouter, and its robot team won two championships at CVPR and ICRA 2026 workshops.
Shanghai's commerce authorities have issued a plan to promote overseas use of computing power and foundation models, and support exports of AI products such as intelligent agents and multimodal generation models.
Global markets were mixed as oil prices surged, chip stocks slumped, and tech giants made strategic moves. Key developments include new tariffs, IPO pricing, and AI model advancements.
Galaxea's open-source VLA model G0.5 ranks first on RoboDojo with an average score of 20.23, including 28.25 in precision and 44.12 in long-horizon tasks.