Unix AI: Rising From Robotics Competitions to Real-World Applications
A Surge in Interest Post-Competition
Since the conclusion of the World Humanoid Robot Games, Unix AI, a key player in robotics, has experienced an unprecedented surge in inquiries. “In the second week after the competition, more than a dozen hotel clients visited us,” shared Yang Fengyu, the visionary founder and CEO of Unix AI.
Medals that Speak Volumes
Back in August, Unix AI made its mark by securing two gold medals and one silver in the hotel cleaning and guest reception categories. This remarkable achievement quickly caught the attention of industry players ranging from hotels to retirement communities, all eager to explore the potential of service robotics.
Robotics Tested: Generalization, Dexterity, and Speed
The competitions at the World Humanoid Robot Games put robots through rigorous testing in generalization, dexterity, and speed. In the cleaning challenge, robots were tasked with rapidly collecting scattered items, while the reception task involved transporting a guest’s suitcase to a specified location. Such tasks highlighted the robust capabilities of Unix AI’s technology.
Learning on the Job: A Proven Strategy
Unix AI’s impressive performance can be attributed to its robots gaining real-world experience in “quasi-consumer” cleaning scenarios. Specifically deployed in hotels, these robots have honed skills in tidying, cleaning, and trash collection by learning on the job. “While they may still lag behind human staff in speed, their ability to operate independently in private settings adds considerable value,” Yang explained.
Future Prospects: Beyond the Hotel Sector
Yang envisions that the skills acquired in hotel environments will transcend to other sectors, including households, restaurants, and even fast-food outlets. The company is already on the path of small-scale deliveries, having forged contracts with hotel chains, property managers, and retirement communities.
A Different Approach to Robotics Development
Contrary to the current trend among robotics startups favoring the end-to-end VLA (vision-language-action) approach, Unix AI adopts a different strategy. Citing the issue of limited training data, the company dissects tasks into key motions and trajectories, implementing imitation learning to enable robots to master new skills with minimal demonstrations. This allows for a more streamlined process of real-world learning.
Yang Fengyu: A Young Visionary
Born in 2000, Yang’s educational journey led him from the University of Michigan to a PhD program at Yale University, which he paused in 2024 to kickstart Unix AI. “With Chinese companies dominating hardware for decades, I recognized a remarkable opportunity,” he remarked.
The Anticipation of the Third-Generation Robot
In a recent discussion with 36Kr, Yang provided insight into the eagerly awaited third generation of Unix AI’s humanoid robot, Wanda. This new model promises to enhance the existing capabilities significantly.
Lessons Learned from Competition
When asked about the aftermath of the competition, Yang noted, “As soon as the event wrapped, our hotline lit up. More than ten hotel clients reached out just in the second week.” He acknowledged that while the competition didn’t attract much attention at the venue, the results sparked interest among potential clients.
Adapting to Challenges with Resilience
The competition also served as a testing ground, pushing the boundaries of Unix AI’s robotics capabilities. The reception event posed a significant challenge, requiring robots to lift and maneuver luggage despite unforeseen hardware issues during execution. The iterative redesign process transformed these challenges into valuable learning experiences.
The Choice of Hotels as a Launchpad
“Hotel cleaning tasks are our starting point for quasi-consumer skills,” Yang explained. Once the robots demonstrate proficiency in such tasks, these skills can catalog and be easily exported to other environments like homes or restaurants. Yang highlighted how hotels provide a wealth of data due to fewer confidentiality restrictions, thereby allowing robots to learn and improve continually.
Experience in Competition as a Benchmark
“The awards we received represent years of accumulated experience, rather than a fleeting preparation,” Yang pointed out. Robots were already well-versed in handling tasks typical of hotel rooms, which included clearing bottles and boxes and disposing of trash — tasks they excelled at.
The Power of Data and Deployment
Unix AI’s process integrates real-world deployment with continual data collection to refine its robotics models. Yang likened their approach to that of Tesla, emphasizing the importance of widely deploying robots to gather valuable real-world data to enhance learning models.
Closing the Door Challenge and Quick Adaptation
The challenge of “closing a door” presented another complex scenario during the competition. Yang explained that the size of the door exceeded their training parameters, which initially disrupted their dual-arm strategy. Remarkably, the team utilized virtual reality (VR) to gather new data and retrain their robots swiftly, showcasing the flexibility and adaptability of their systems.
Understanding UniFlex and Imitation Learning
At the heart of Unix AI’s approach is UniFlex, a perception-operation decoupled model designed for key-point imitation learning. “With just a handful of demonstrations, our robots can generalize learned tasks,” Yang stated, underscoring the flexibility of their systems in adapting to varied scenarios.
Why Not VLA? A Tactical Decision
While acknowledging the long-term potential of VLA approaches, Yang emphasized that current data limitations make them impractical for immediate application. Unix AI’s focus remains on efficiency and practicality rather than the latest trends in robotics development.
VTLA and Tactile Input: The Future of Interaction
Yang discussed the importance of tactile sensing, known as VTLA, and how Unix AI’s UniTouch system integrates vision and tactile data for improved interaction. This innovative approach allows robots to visualize object handling based solely on appearance, enhancing their operational capabilities.
Hardware Stability and Full-Stack Development
For Unix AI, hardware development is pivotal. Yang noted that they prioritize full-stack hardware design for three key reasons: control over production, cost reduction by eliminating middlemen, and ensuring data consistency across generations of robots.
A Standout Product: The Third-Generation Wanda
The latest iteration of the Wanda robot symbolizes a shift in design focus. “The third generation is tailored for efficiency and practical work,” Yang explained, highlighting its simplicity and power over human-like aesthetics.
An Innovator’s Perspective in Robotics
Yang’s youth doesn’t limit him; instead, he views it as a competitive advantage. “Young innovators aren’t shackled by outdated paradigms,” he said, indicating that fresh ideas are critical to driving forward the realm of embodied intelligence.
Conclusion
Unix AI is at the forefront of robotics innovation, combining practical experience with cutting-edge technology. As it expands its horizons from hotel settings to broader applications, the company’s unique strategies and grounded vision position it as a promising player in the evolving landscape of service robotics. With continuous advancements, such as the third-generation Wanda robot, Unix AI is poised to redefine industry standards and expectations.
