A Chinese robotics startup has demonstrated humanoid robots and fixed dual-arm robotic systems working together on an end-to-end logistics task using a single AI model.

The system by MindOn, called Mind-0, enables different robot types to coordinate under a unified intelligence framework trained entirely on human-centric data rather than robot-collected task data.

To support multiple hardware platforms, MindOn separates high-level reasoning from low-level motion control, allowing the same AI brain to operate across different embodiments.

The Shenzhen-based company says its execution compensation model delivers sub-centimeter manipulation accuracy in real-world environments.

MindOn’s new robotics framework enables different types of robots to operate under a single AI system, marking a step toward hardware-agnostic embodied intelligence.

At the core of the approach is Mind-0, a unified model that separates high-level reasoning from low-level motion control. The architecture allows one intelligence layer to handle perception, task planning, and decision-making, while specialized controllers translate those decisions into movements suited to each robot’s physical design. This enables the same AI model to be deployed across multiple robotic platforms without requiring separate training pipelines, reports Humanoids Daily (HD).

A key feature of the system is its reliance on human-centric training data rather than robot teleoperation. Instead of collecting demonstrations by manually controlling robots, MindOn trains its models using human motion captured through whole-body tracking systems, egocentric cameras, and handheld devices. The company argues that this approach preserves natural human problem-solving behavior and avoids the limitations often introduced when operators adapt their actions to a robot’s constraints.

To bridge the gap between human demonstrations and robotic execution, MindOn developed a cross-embodiment data pipeline that converts human actions into representations usable by different robotic systems. Supporting this process is a whole-body action model trained on large-scale motion capture datasets, enabling robots to maintain balance, coordinate movements, and execute tasks while respecting their own physical limitations, reports HD.

Universal robot control

The company has also introduced a Real-World Execution Compensation Model to address the long-standing sim-to-real challenge. Models trained in simulation often experience performance degradation when deployed on physical hardware due to differences in dynamics and environmental conditions. MindOn says its compensation system uses a small amount of real-world deployment data to correct these discrepancies, allowing robots to improve tracking accuracy and manipulation performance. According to the company, the technology achieves sub-centimeter manipulation precision on the Unitree G1 humanoid platform.

Another component of the framework is a hierarchical reasoning system designed to account for execution delays in physical robots. Human demonstrations are naturally free of mechanical latency, but robots must cope with sensing, computation, and actuation delays. MindOn’s system continuously monitors feedback from low-level controllers and adjusts command timing in real time to maintain synchronization between planning and execution.

The company recently demonstrated the technology using a mixed fleet consisting of Unitree G1 humanoid robots and stationary dual-arm robotic systems. In the demonstration, the robots collaborated on a logistics workflow involving item retrieval, transportation, packing, and box sealing. Despite their different physical configurations and capabilities, all robots operated using the same underlying AI model.

Founded in Shenzhen in 2025, MindOne Robotics is positioning its technology as a scalable alternative to teleoperation-heavy robotics development. The company plans to expand its human-centric datasets and extend deployment to additional robotic form factors, including mobile dual-arm systems, as it pursues the goal of a universal robotic intelligence capable of operating across diverse hardware platforms.

Recommended Articles