From Models to Productivity: Galbot and Its Industry Partners Explore the Next Frontier of Embodied Intelligence
Making embodied intelligence a reality across industries.
On August 21, during the 2026 World Robot Conference, Galbot hosted the “From Models to Productivity” Technology and Industry Applications Forum at the Beiren E-Creative Exhibition Center. The forum brought together leading researchers from universities and institutions in China and abroad, developers, and business executives to discuss innovations in embodied foundation models, robot evolution, commercialization, and the boundaries of technology. Together, they presented the latest thinking on technological advances and industrial deployment.
Three Major Advances in the Technology Foundation, with G0.5 MAX Coming Soon
In his speech, “From Models to Productivity: Building a Closed-Loop System for General-Purpose Embodied Intelligence,” Galbot Chief Scientist Zhao Xing noted that embodied intelligence is evolving through three stages: instinctive intelligence, task-oriented intelligence, and evolutionary intelligence. The industry is moving beyond breakthroughs in isolated technologies and entering a new phase of systemic competition encompassing models, data, robot bodies, infrastructure, and applications. Based on this view, Galbot has built a complete productivity flywheel for embodied intelligence: starting from industrial scenarios to accumulate real-robot data, using data to drive iterative model evolution, enabling robot bodies to perform physical tasks through the models, relying on infrastructure to support the continuous optimization of robot fleets, and ultimately feeding the results back into industrial applications to form a self-reinforcing positive cycle.
The technology foundation has seen three major advances:
First, the G0.5 foundation model has been upgraded, alongside an open-source development initiative.
The G0.5 foundation model innovatively incorporates an intrinsic chain of thought for action, unifying visual perception, language understanding, logical reasoning, and action generation within a single autoregressive data stream. Since its release, G0.5 has comprehensively surpassed the industry SOTA across eight internationally recognized benchmarks, including LIBERO and RoboTwin 2.0. The G0.5 MAX version is also coming soon, with further enhancements to the model’s overall capabilities.
To advance the industry’s open-source ecosystem, Galbot has officially launched the G0.5 Reproduction Program. The program is open to research institutions and developers worldwide, providing full access to model weights, inference interfaces, evaluation benchmarks, and a complete suite of fine-tuning tools. Developers are encouraged to conduct in-depth validation in three areas: “simulation reproduction,” “real-robot reproduction,” and “open-source co-creation.”
Second, Galbot previewed its pre-trained foundation model program based on the Fast-WAM world model.
During inference, this model skips the step of generating future video. It reduces single-step inference latency from approximately 800 milliseconds in conventional paradigms to 190 milliseconds, delivering more than a fourfold speedup. Zhao Xing revealed that a pre-trained foundation model based on the Fast-WAM architecture will also be officially released, offering faster inference and stronger generalization across tasks, objects, and scenarios.
Third, the G-Fleet distributed reinforcement learning system was officially released. Designed as distributed infrastructure for the continuous learning of real-robot fleets, G-Fleet fully supports the entire process of fleet-level policy deployment, parallel real-robot Rollout, real-world data feedback, human intervention and calibration, distributed reinforcement learning, automated evaluation, and version release. It closes the loop of “model deployment → real-world interaction → data filtering → iterative training → new version release,” addressing a major industry challenge: how to enable large fleets of deployed robots to evolve continuously, safely, and efficiently in physical environments.
Led by the Next-Generation Wheeled Dual-Arm Robot Nexo, Three New Robot Platforms Bring Productivity to Life
Galbot Partner and COO Li Tianwei delivered a speech titled “The Evolution of Next-Generation Robot Platforms for Embodied Intelligence Productivity.” At the outset, he identified the central challenge in robot platform design: “For intelligence, iteration is endless; for hardware, however, focus is essential.”
He broke down the idea that “intelligence defines the robot body” into three engineering principles: intelligence prioritizes speed, so the robot must provide the required electromechanical cycle time; intelligence prioritizes precision, so the robot must deliver 0.1 mm repeatability and 0.1 N force-control accuracy; and intelligence prioritizes generalization, so the robot must support complete transfer from teleoperation data to Egocentric data. A robot platform is not merely a passive “shell,” but a physical extension of intelligent capabilities.
“Today, intelligence is no longer the only force defining the robot body. We now have a new group of influential voices: industry customers.” Li Tianwei noted that customers are demanding levels of reliability, service life, and adaptability to alternating hot and cold environments far beyond those expected of laboratory prototypes. These have been key areas of focus for Galbot over the past year. The latest developments across three major robot platforms were also announced at the event:
First, the next-generation flagship wheeled-arm humanoid robot Nexo, positioned as an “embodied platform” for general-purpose tasks. The robot features force and tactile capabilities powered by cloud-based control across the entire system, with a peak dual-arm payload of 20 kg. It is designed for four major application scenarios: e-commerce and retail, industrial manufacturing, express delivery and logistics, and commercial services. Nexo offers eight hours of battery life, supports continuous 7×24-hour operation, and provides three power replenishment modes: fast charging, battery swapping, and autonomous recharging.
Second, the bipedal humanoid robot Kengo. Its electromechanical transmission system is entirely developed in-house by Galbot, targeting unstructured scenarios that are difficult for wheeled platforms to access, including energy, power, large-scale exploration, and equipment inspection.
Third, Lemo, designed for the developer ecosystem. Lemo is a cost-effective dual-arm desktop manipulation platform capable of highly delicate tasks, such as threading a needle, through its self-developed teleoperation system.
Li Tianwei also noted: To date, Galbot has delivered orders for thousands of model-driven robots operating in real production environments, with a delivery target of more than 10,000 units next year. This achievement has been made possible by capacity-building efforts with industry partners, as well as continuous iteration and testing of the robot platforms in terms of reliability, service life, and adaptability across a wide temperature range. “The success of embodied intelligence will not belong to any single company. It will require the collective efforts of the entire industry ecosystem.”
Diverse Global Perspectives Converge as the Next Generation of AI Innovators Exchange Ideas
Two roundtable discussions explored embodied intelligence from the perspectives of commercial deployment and technological advancement.
The Industry Roundtable was moderated by Yu Lei, Partner and President of Marketing and Services at Galbot. Participants included Ashish Kapoor, Founder and CEO of General Robotics; Edwin Yap, Founder of Mr. Robot and CEO of Jemco Group; Wang Kai, Chief Technology Officer of Demae Technology Group; and Tian Ruilin, Senior Embodied Intelligence Expert at JD Retail.
Yu Lei noted that whether robots can truly deliver productivity depends on multiple factors, including whether market demand, technological iteration, data accumulation, the supply chain, and business models can be effectively integrated. However, he believes that day will arrive sooner than expected. Wang Kai stated candidly that overseas markets offer shorter investment payback periods and that customers there are more accepting of robots’ current capabilities. Demae Technology’s strategy is therefore to “go overseas first, then deepen its presence in China.” He also pointed out that whether ROI can ultimately be justified remains the key bottleneck to large-scale deployment. From the perspective of a major technology company, Tian Ruilin added that JD plans to deploy robots at a scale of one million by 2028, using a “general-purpose foundation + vertical models” architecture to empower industries across the board.
International guests offered insights from different markets. Ashish Kapoor said that North American companies already regard robots as a “must-answer question,” but overly high ROI expectations and high employee turnover remain long-term obstacles to deployment. Drawing on experience in Southeast Asia, Edwin Yap noted that the region cannot adopt a one-size-fits-all strategy. Markets such as Malaysia and Vietnam have lower costs and numerous manual-operation scenarios, making them natural environments for collecting robot data. Multiple speakers noted that cross-company and cross-border ecosystem collaboration will become an industry trend.
The Technology Roundtable was moderated by Chen Qian, Co-Founder of Silicon Valley 101. Participants included Zhao Xing, Chief Scientist of Galbot; Yang Shiyuan, Co-Founder of DynaRobotics (online); Huang Siyuan, Director of the Embodied Robotics Center at the Beijing Institute for General Artificial Intelligence; Yu Chao, Assistant Professor at the Institute of Artificial Intelligence and Robotics, Tsinghua Shenzhen International Graduate School; Zhu Yichen, Founder and CEO of Current Robotics (Yuanliu Intelligence); and Li Haoxuan, Assistant Professor at the Institute for Artificial Intelligence at Peking University.
Addressing the debate over whether world models are merely an industry buzzword, the panelists agreed that technical approaches in the industry have yet to converge. Some teams are focusing on representation learning, others on simulator development, while still others are turning to 3D modeling.
Yang Shiyuan shared counterintuitive data from Silicon Valley: after pre-training on millions of hours of human video, model accuracy on robot test sets actually declined, highlighting the enormous gap between “understanding the world” and “taking action.” Zhao Xing, meanwhile, argued that the potential for collecting embodied intelligence data is sufficiently vast, and that as industrial adoption advances, there is no need to be overly concerned about data supply. Yu Chao pointed out that pre-training defines the capability boundary, while post-training is the key to expanding it. However, post-training currently relies heavily on human intervention, and large-scale implementation remains some distance away. From the perspective of causal inference, Li Haoxuan added that existing training largely remains at the level of correlation. Robots need counterfactual reasoning capabilities to truly understand the causal relationship between actions and outcomes.
At the conclusion of the roundtable, each guest summarized the industry’s most important area for advancement in a single sentence. Their views converged on five key priorities: native multimodal model architectures, efficient use of human data, causal reasoning capabilities, open-source development of foundational infrastructure, and closed-loop iteration in high-value scenarios.
The forum showcased Galbot’s comprehensive capabilities across models, infrastructure, and robot platforms. Embodied intelligence is currently at a critical stage, transitioning from laboratory prototypes to industrial productivity. This transition requires not only continued breakthroughs in foundational algorithms, but also hardware engineering refinement, open-source ecosystem development, and the accumulation of experience in real-world scenarios. Going forward, Galbot will continue to drive the coordinated evolution of models and robot bodies, working with partners across industry, academia, and research to bring embodied intelligence into real-world use across industries.