MindOn Robotics: Pioneering the Future of Humanoid Household Robots

From Tencent Robotics X to Founding MindOn Robotics

Zhu Qingxu, a robotics researcher born after 1995 and formerly part of Tencent Robotics X, launched MindOn Robotics in June. The company specializes in embodied intelligence with a strong emphasis on algorithm development. Within just four months, MindOn successfully secured nearly RMB 100 million (approximately USD 14 million) through three funding rounds:

• The initial seed round was exclusively funded by Oriza Seed.
• The second round was spearheaded by MSA Capital, with additional investments from Innoangel Fund, Yuanbio Venture Capital, Genspark, and Oriza Seed, which increased its stake.
• The third round saw Genspark take the lead, joined by Plum Ventures, G&O, Innoangel Fund, and continued participation from MSA Capital.

Demonstrating Advanced Household Robotics Capabilities

Recently, MindOn integrated its proprietary algorithms with Unitree Robotics’ hardware, showcasing robots performing complex household chores with near-human dexterity. Demonstrations included a robot watering plants while balancing on a greenhouse frame and meticulously removing dust mites from bedding-tasks executed in real-time without video acceleration.

These scenarios were inspired by a social media post depicting the demanding daily routine of a single mother. Zhu selected tasks that required intricate coordination of limbs, highlighting the challenges of embodied intelligence. The videos quickly gained widespread attention online.

Zhu’s Vision: Humanoid Robots as the Optimal Household Solution

Contrary to the prevailing industry consensus, Zhu believes that bipedal humanoid robots are the most suitable for domestic environments and predicts their practical deployment within three to five years. While many experts anticipate a decade-long timeline due to the complexities of motion control, balance, and diverse home layouts, Zhu argues that the human form offers unmatched adaptability.

He emphasizes that human-like robots can leverage extensive human motion data, enabling them to navigate stairs, uneven surfaces, and confined spaces more effectively than wheeled robots. Movements such as climbing, crouching, and stepping on objects are inherently easier for humanoid designs.

Challenging the Teleoperation Paradigm

Zhu critiques the widely used teleoperation method, where operators manually control robots via handheld devices. This approach relies heavily on conscious thought, resulting in slow, unnatural, and jerky robot movements. He contends that training robots on such data limits their fluidity and efficiency.

Drawing from his academic background-graduating from a joint ETH Zurich and RWTH Aachen program in 2021-and four years of industry experience at Robotics X, Zhu and his team discovered that teleoperation-based models underperform in execution speed and naturalness.

Supporting Zhu’s stance, Boston Dynamics recently highlighted similar concerns, noting that teleoperation’s dependence on conscious human control hampers dynamic and efficient robot behavior.

Innovative Data Collection and Training Architecture

MindOn employs a dual-structure model inspired by human neuroanatomy: a “cerebellum” responsible for motion control and a “cortex” dedicated to planning and generalization. Currently, the company prioritizes enhancing the cerebellum, an area often overlooked in robotics.

To gather high-quality training data, MindOn uses optical motion capture technology combined with a Universal Manipulation Interface (UMI). Operators wear motion capture suits and perform natural movements in a controlled environment, capturing subconscious coordination. Subsequently, they use UMI handheld grippers to interact with objects in real-world settings, generating extensive hand-object interaction datasets.

Investor Confidence and Market Outlook

Zhu attributes MindOn’s rapid fundraising success to its distinctive technological approach, which complements existing portfolios of investors already active in embodied intelligence. He forecasts that humanoid robots will enter retail and fast-casual dining sectors-such as unmanned fast-food outlets-within one to two years, where controlled environments and repetitive tasks facilitate swift deployment.

He envisions household adoption within three to five years as training efficiency improves, with robots arriving pre-equipped with essential skills and later learning new tasks through simple user demonstrations.

Addressing Industry Challenges and MindOn’s Competitive Edge

Zhu identifies generalization-adapting learned skills to diverse objects, positions, and environments-as the primary hurdle for household robotics. He stresses the importance of large-scale, high-quality data to overcome this challenge, aligning with scaling laws in machine learning.

Regarding MindOn’s competitive advantage, Zhu emphasizes the company’s ability to identify fundamental flaws in prevailing methods and iteratively develop innovative solutions. This adaptability and commitment to long-term technical depth form their true moat.

Insights from Zhu Qingxu: A Q&A Overview

Why is teleoperation fundamentally flawed?

Zhu explains that teleoperation forces operators to use deliberate, conscious control, which is inherently slow and unnatural. This results in jerky robot movements and limits performance. Tasks requiring fine tactile feedback, like twisting a bottle cap, suffer further due to lack of force sensation.

Why did teleoperation become widespread despite its drawbacks?

It was the first practical method to collect real-world manipulation data, enabling initial progress in robot control.

How does MindOn’s data collection system improve upon teleoperation?

By combining motion capture of instinctive full-body movements with UMI-enabled object manipulation, MindOn captures high-fidelity, subconscious human actions at scale, surpassing video-only or teleoperation methods.

Can you describe MindOn’s training architecture?

The “cerebellum” learns fundamental human motions like walking and grasping from motion capture data, forming a universal skill library. The “cortex” manages perception and planning, orchestrating these skills to perform complex tasks.

How will humanoid robots function in unmanned retail environments?

Robots learn all necessary actions in the lab, then adapt to real stores using UMI data. This process is efficient, requiring only a few days to master all roles, unlike teleoperation.

How is scenario-specific training handled without motion capture studios in homes?

Basic movements are captured in the lab, while environment-specific data is collected on-site, enabling robots to generalize to household tasks.

What challenges exist in generalizing robot skills to real homes?

Robots must generalize across different objects, positions, and scenes, necessitating diverse and extensive datasets.

Why are humanoid robots preferred for household use?

Homes are designed for humans, with stairs, uneven floors, and furniture layouts favoring bipedal robots. Human motion data is abundant, unlike for alternative robot forms.

How will users teach new tasks to robots?

Robots will come pre-trained with common skills. For new tasks, users may demonstrate actions once, enabling robots to learn through observation and minimal practice.

What is MindOn’s unique advantage in the robotics market?

MindOn’s strength lies in recognizing the limitations of existing methods and iteratively developing novel solutions, turning early-stage ideas into practical technologies.

What is the future outlook for the embodied intelligence sector?

Zhu anticipates industry consolidation, with only companies focused on genuine technical progress thriving. MindOn aims to deepen its foundational capabilities to remain a leader as the market matures.