This post is a collaboration with Google Gemini 2 advanced, where the transcript of the YouTube video Will robots explode in capabilities like LLMs did? (10 January 2025) by Dr Waku was transformed into a blog post/ note.
Artificial intelligence (AI) has been making leaps and bounds lately, especially with the jaw-dropping advancements in Large Language Models (LLMs) like ChatGPT. These models can write poems, answer complex questions, and even generate code, leaving many of us wondering: will robots experience a similar breakthrough?
To explore this question, let’s dive into a fascinating interview by Dr. Waku on his YouTube channel. He chats with David Watkins, a researcher at the AI Institute in Boston, a leading organization dedicated to pushing the boundaries of robotics. The key takeaway? Robotics faces a serious data bottleneck, which is holding it back compared to other AI fields.
The AI Institute, founded by Marc Raibert (the original founder of Boston Dynamics) and funded by Hyundai, is a unique place. It’s a private company, but it operates like a pure research institution, free from the pressures of product development. This allows researchers like David to focus on the long-term challenges that need to be solved to truly unlock the potential of robots.
In this blog post, we’ll unpack the key insights from their conversation, exploring the state of robotics research, the challenges of data collection, and the exciting possibilities that lie ahead.
The State of Robotics Research
The AI Institute’s Approach
Imagine a place where you can tinker with all sorts of robots – from the agile Spot (yes, the famous Boston Dynamics “dog”) to a whole array of robotic arms – without worrying about deadlines or investors. That’s the reality at the AI Institute. This unique environment fosters collaboration with academia and provides researchers with access to a diverse range of platforms. It’s a multidisciplinary playground where hardware engineers, software developers, and machine learning experts come together to tackle the big questions in robotics.
Hardware vs. Software
One of the most surprising revelations from the interview is that, despite the impressive capabilities of today’s robotic hardware, we’re barely scratching the surface of what’s possible. David estimates that current software algorithms are only using about 10% of the hardware’s potential. Think about it – that sleek robotic arm you see in videos? It’s capable of so much more!
This highlights a crucial point: we need a co-design approach. It’s not just about building better robots; it’s about developing software that can truly harness their power.
The Analogy to LLMs
The success of LLMs like ChatGPT can be largely attributed to two factors: massive amounts of training data and the scaling up of model size. These models have been trained on vast portions of the internet, learning from the collective knowledge of humanity. Could a similar “ChatGPT moment” happen for robotics? It’s a tempting thought, but as we’ll see, the path for robots is a bit more complicated.
An Exploratory Phase
It is important to understand that, unlike the field of Natural Language Processing (NLP) with its established paradigms, robotics is still very much in its exploratory phase. As David emphasizes, “Robotics is still in a very exploratory phase.” There aren’t standardized platforms, methodologies, or even a widely agreed-upon vocabulary yet. This makes research exciting but also challenging, as researchers are still figuring out the fundamental building blocks of intelligent robotic systems.
The Data Bottleneck
Why is Data Collection Harder for Robots?
Here’s the crux of the issue: unlike LLMs, which can feast on readily available text data, robots lack a comparable “ground truth” dataset. For language models, human-written text provides a clear example of desired behavior. Robots, on the other hand, often have to learn through trial and error, using a technique called reinforcement learning (RL).
RL is much more sample-inefficient than the imitation learning used in LLMs. Imagine trying to learn how to walk without ever seeing someone else do it – that’s the challenge robots face. On top of this, collecting data in the real world is messy and complex. Robots need to navigate diverse environments, interact with objects, and adapt to unexpected situations.
Further complicating matters is the “state space explosion” problem. The sheer number of possible states and actions a robot can be in is mind-bogglingly large, making it computationally difficult to model every scenario. It is different in comparison to a Sudoku puzzle. In Sudoku you can follow a strategy and if you make a mistake and hit a wall, you cannot proceed. With robots however, you can make corrections more easily, although it is still possible to fail.
Current Approaches to Data Collection
Despite these challenges, researchers are exploring innovative ways to gather the data robots need:
- Teleoperation: One promising approach is teleoperation, where humans control robots remotely to demonstrate desired behaviors. Chelsea Finn’s work at Stanford with Mobile ALOHA is a great example of this. However, directly translating human actions to robots is not straightforward. Our hands are vastly different from robotic grippers, making it challenging to map human demonstrations onto robot control policies. The “Universal Manipulation Interface” paper is also connected to this problem.
- “Mechanical Turk” for Robots: Companies like Sensei Technologies are building platforms that connect researchers with a network of human operators who can collect data remotely – a kind of “Mechanical Turk” for robotics. They want to “scale up data collection for robotics” says David.
- The Limitations of “Arm Farms”: While the idea of setting up a lab full of robots continuously performing tasks (like the Google arm farm) sounds appealing, it turns out that this approach has limitations. The lack of environmental diversity restricts the generalizability of the learned skills. As David puts it, it’s like teaching a kid to play only one video game and then expecting them to master all others.
- The Potential of Simulation: Another promising avenue is the use of simulation to train robots. Simulation allows researchers to create virtual environments and generate large datasets without the constraints of the physical world. For instance, work from ETH Zurich has shown that it’s possible to train quadruped robots to walk purely in simulation. However, there’s still the “sim-to-real” gap – the challenge of transferring skills learned in simulation to real-world robots.
Many of the arms used at the AI Institute utilize impedance-based control, which allows the robot to regulate its stiffness and interact with objects without shutting down due to safety limits when it detects an unexpected force.
The Role of Human Feedback
Just as Reinforcement Learning from Human Feedback (RLHF) played a crucial role in refining ChatGPT’s responses, human feedback can also guide robot learning. Even simple signals like a “thumbs up” or “thumbs down” can provide valuable information about whether a robot’s actions are desirable. Humans possess an intuitive understanding of physical interactions, making their feedback a powerful tool for shaping robot behavior.
The Future of Robotics
The Importance of Embodiment
David strongly believes that embodiment – having a physical presence in the world – is essential for achieving higher levels of AI intelligence. Robots, unlike disembodied AI systems, receive direct feedback from their actions. This physical interaction is crucial for developing a true understanding of the world and overcoming the biases that can plague AI systems trained solely on abstract data. There could also be implications for AI safety. It might be easier to align AI if it is embodied in robots.
Active Learning and Curiosity
Active learning, where AI systems can proactively seek out the data they need, holds great promise for making data collection more efficient. However, implementing active learning in robotics is tricky. It requires a shared vocabulary and understanding between humans and robots, something we’re still working towards. It would enable robots to say “That thing – more of it!”
The Need for Innovation
Simply scaling up existing methods might not be enough to unlock the full potential of robotics. We need new research directions, particularly in developing algorithms that can learn from multimodal data (combining information from different sensors) and effectively incorporate human feedback. David emphasizes that it’s not just about bigger models or better hardware; it’s about fundamentally rethinking how robots learn.
Call to Action
The world of robotics is brimming with exciting possibilities, and there are many ways to get involved. Check out the resources mentioned in the video, such as:
- The AI Institute: https://theaiinstitute.com/
- Universal Manipulation Interface paper: https://arxiv.org/abs/2402.10329
- Sensei Technologies: https://www.ycombinator.com/companies/sensei
- Chelsea Finn’s work on Mobile ALOHA: https://mobile-aloha.github.io/
- Data Scaling Laws in Imitation Learning for Robotic Manipulation paper: https://arxiv.org/abs/2410.18647
- RoboCrowd paper: https://arxiv.org/abs/2411.01915
- Ego4D paper: [invalid URL removed]
The AI Institute is also hiring, so if you’re passionate about robotics, reach out!
Nowadays, robotics has become much more accessible for hobbyists and enthusiasts, thanks to 3D printing and cheaper components. So, if you’re curious, why not explore some online tutorials, open-source robotics projects, or affordable robot kits? You might be surprised at what you can build! For example, you could take a look at David’s PhD thesis. He describes a robot that gets placed in an environment and is told to pick up an object. This could be an inspiration for your own project.
Conclusion
The data bottleneck is a significant hurdle in the race to build truly intelligent robots. But with researchers like David Watkins and organizations like the AI Institute pushing the boundaries of what’s possible, the future of robotics looks incredibly bright. While we may not have a “ChatGPT moment” for robots just yet, the ongoing advancements in data collection, coupled with a deeper understanding of how robots learn, promise to revolutionize the field in the years to come.
The relationship between humans and robots is poised to become increasingly intertwined. As robots become more capable and integrated into our lives, it’s essential to consider the ethical implications of this technology. How can we ensure that robots are developed and used responsibly? What role will they play in our society? These are questions we must grapple with as we move forward.
Leave a Reply