Candid Lessons in Building Humanoids

A few years ago, at a technology event, a senior executive from a leading Japanese robotics company pulled me during the networking session and shared a candid insight with me. He said, "You are not learning from our failures in building humanoids. Humanoids are not the way to go." Initially, I disagreed with him, but as time passed and after years of designing robots and earning a PhD in the field, I realized the depth of his words. Today, as a new wave of humanoids emerges and outsiders share thrilling demo videos, I feel compelled to share the lessons learned.

While I am excited about building robots, the recent spate of demonstrations from Musk, Jensen and others has the risk of making the same mistakes.

1. The Demo Problem

No industry showcases demos as exciting as the robotics industry. Over the decades, we have witnessed incredible feats: Honda Asimo descending stairs and playing soccer, Willow Garage's PR2 making breakfast, Boston Dynamics' Atlas performing backflips and dancing, and Hanson Robotics' Sophia engaging in seamless conversation and even gaining citizenship. These are just a few examples among hundreds.

However, years later, none of these robots have been commercialized at a large scale. Rather than benefiting the industry, such demos can harm it by setting public expectations that are unattainably high. The general public now has an inflated perception of robotic capabilities, leading to disappointment when real-world applications fall short.

2. Vanity vs. Sanity Problem

At Mitra Robot, we developed an internal metric, ranging from 0 to 1, to evaluate customer requests where 0 represents 'Vanity' and 1 represents 'Sanity'. The robotics industry has witnessed more failed ventures than perhaps any other, often due to the vanity over sanity approach. High-profile individuals like Jensen and Musk are among many who have faced this challenge, though they had other successful ventures to fall back on.

The allure of creating humanoid robots often distracts from practical and immediately useful applications, such as automating mundane tasks like cleaning or logistics in industrial settings. Interestingly, the evolution of modern Large Language Models (LLMs) from the exciting phase of ChatGPT to more routine tasks like document management through Retrieval-Augmented Generation (RAG) systems exemplifies a shift from vanity to sanity, leading to more sustainable success.

3. Fundamental Hardware Challenges

The journey of companies like Boston Dynamics epitomizes the hardware challenges in robotics. Marc Raibert had an excellent team 32 years ago when he branched out from the MIT's Leg Lab building dynamic robots. Boston Dynamics excelled in brilliant robots but faced commercial hurdles. 21 years later, Google X under Andy Rubin, believing that more funding and intelligence could resolve these issues, acquired Boston Dynamics but eventually realized the complexity of the challenge. It is not a question of intelligence, money or Google's market access that could make a dent in this.

Years later Softbank under Mayoshi Son thought it was more of a vision problem and brought in Boston Dyanmics and also acquiring Aldebaran and investing in humanoid research such as the Romeo robot.

Now, Boston Dyanmics has shifted to more mundane industrial monitoring problem through Spot after getting acquired by Hyundai in 2020 and Softbank Robotics shifted focus towards more practical applications like and cleaning robots.

4. Unnecessary complexity

When we designed the Mitra, we went for a semi-humanoid design keep the bottom wheeled, but arms that mimics human arms. Even with this the problem were enormous. For starters, customers were asking the arms to get shorter as the arms were too long [like a monkey]. However, if we shorten the arms, there was no space to fit the motors and cooling elements. The only reason people wanted it shorter was to unnaturally replicate human proportions. We went for simpler solution of keeping the elbows bent all the time so that the arm doesn't look long. These kind of finer engineering distracts from the bigger problem and solution. Most robotics companies get submerged from these.

We thus decided to focus on non-android robots, where we have a lot of flexibility to place the sensors and actuators. However, the challenge is not merely HRI based.

Despite advancements in computational power, fundamental hardware limitations, particularly in materials, battery life, and motor technology, continue to hinder the development of efficient biped humanoids. Softbank Robotics attempted to overcome these through acquisitions, like Schaft Robotics, which developed innovative solutions for energy storage and motor cooling.

Yet, these innovations were not sufficient to conquer the challenges of creating a commercially viable humanoid robot. It is not the smartness or lack of AI that killed these products, but more fundamental hardware challenges that would allow a day of operation under a single charge. If you had seen Boston Dynamics demo booths in any event, you might notice a number of spot dogs resting below. Even the more efficient quadrapeds such as the Spot dog cannot run longer than an hour.

Wheels, not legs, revolutionized mobility because of their efficiency and simplicity. However, the allure of bipedal robots continues to captivate the imagination of the public and investors, despite the practical advantages of wheeled robots.

Conclusion

This is an exciting era for robotics, but it is crucial to heed the lessons from decades of trials and errors in the industry. Each new venture optimistically claims, "This time it is different," yet the history of evolution teaches us that progress is a result of cumulative learning and adaptation over millennia. We cannot hope to compress this vast experience into a single lifetime or project, but we can learn from it to make informed and pragmatic decisions in the development of robotics.

What has worked in robotics is more mundane, boring problems in the DDD [Dull, Dirty, Dangerous] category.