Mind, Motion, and Machines: Key Challenges in Modern Robotics

Introduction

Robots are making their mark on our world, from automated arms in factory floors to robotic vacuums cleaning our living rooms. As technology continues to advance, these intelligent machines are becoming increasingly capable of performing tasks once thought to be exclusively for humans. But still, we do not see robots Sci-Fi would have you believe that we should have robots walking around and working alongside us already. So where are they? Truth is that, despite the remarkable advancements in robotics, several significant challenges remain to be overcome before robots can truly achieve full autonomy and effectiveness. Addressing these obstacles is crucial for ensuring that these machines can safely and reliably integrate into our daily lives. We highlight some of these in this article.

Perception

Robots face major perception challenges in dynamically changing environments. They need to accurately identify and categorize objects in varying conditions, including fluctuating lighting, clutter, and motion. Perception also involves interpreting complex scenes, understanding spatial relationships, and distinguishing between different objects in a way that is reliable enough to support responsive actions. This requires integrating multiple sensory inputs—like vision and depth—to form a cohesive understanding of the environment.

For robots to work alongside humans, they need to recognize, interpret, and sometimes predict human actions and intent. Variability in human behavior, body language, and gestures make perception and interaction complex. Real-time human recognition is resource-intensive and challenging, especially in cases where multiple people are present or are moving unpredictably. Besides, fluid and effective human-robot communication is desirable, which can only be achieved by using natural language.

Flexibility and Adaptability

While impressive in their precision and efficiency, robots often struggle with flexibility, a crucial aspect of real-world applications. This challenge stems from several factors: First, the programming of robots typically requires specific instructions for each task, limiting their ability to adapt to unforeseen changes or variations. Second, the hardware limitations of robots, such as joint stiffness or limited sensory capabilities, can hinder their ability to perform tasks that require dexterity or adaptability. Third, the complexity of real-world environments, with their unpredictable factors like varying lighting conditions, object occlusions, and human interactions, can make it difficult for robots to maintain flexibility and robustness.

While artificial intelligence has notably progressed, current AI models often struggle with the flexibility required for robots to seamlessly navigate new tasks and environments. This limitation stems from the way these models are typically trained: on vast datasets of specific tasks, they excel at those particular functions but may falter when faced with unfamiliar situations.

One of the most used techniques for teaching robots new skills is Reinforcement learning (RL), which lacks of enough flexibility. RL agents can overfit which means that they become too specialized in their training environments. They might excel at a specific task but struggle to adapt to new or slightly different scenarios. This limits their usefulness in real-world applications where unpredictability is the norm.

Intelligence and Cognition

Although robots can perform some tasks autonomously, they lack general intelligence that can enable them to solve complex problems or react to new challenges or unexpected situations. This lack of general intelligence can be a big barrier that prevents robots from performing tasks that humans do or effectively cooperate with them.

Generative AI is a big promise in this regard. Large language models (LLMs) can turn human instructions into robot plans. But what if the robot needs to adapt to unexpected situations? Large vision language models (LVLMs) take things a step further by incorporating visual information into their plan generation. This means robots can not only understand commands but also use their surroundings to make smarter decisions.

Vision language action (VLA) models are the next frontier in generative AI for robotics. These models can represent the physical world in a way that allows robots to understand and interact with it. While VLA models hold immense promise, they require massive amounts of training data and can suffer from hallucinations, similar to LLMs. This poses significant safety concerns when working with physical agents.

Dexterity and Manipulation

One of the most challenging tasks for robots is to manipulate objects dexterously. Grasping an object is a deceptively complex task. It involves several key challenges like accurately measuring the distance to the object, choosing the optimal point to grasp an object -- which depends on its shape, size, and orientation -- and applying the right amount of pressure is crucial for preventing objects from slipping or being crushed. Too much or too little force can result in a failed grasp.

Reinforcement learning techniques is one of the main approaches to teach dexterity to robots. However, there are several challenges to be addressed. Many of these tasks result in rare success when training the model which makes it difficult for the robot to learn from past experiences. Also, in this task the reward is typically delayed since the usefulness of a specific attempt can only be evaluated after the robot precisely accomplishes the task which makes training more difficult.

Imitation learning has emerged as a powerful technique for imparting dexterous skills to robots. A key enabler of imitation learning is the availability of responsive and precise teleoperation methods. These methods allow human operators to remotely control robots, providing them with the ability to execute complex tasks in real-time and generating valuable data for training machine learning models. Imitation learning relies on a specialized form of reinforcement learning. In this approach, robots are rewarded based on how closely their actions match those of the human teleoperator. By optimizing their behavior to maximize rewards, robots can gradually refine their skills and become more proficient at imitating human actions.

Data, Data, Data

Finally, one of the biggest issue with robotics is lack of data. Modern artificial intelligence techniques require vast amounts of data to train models; this is what made LLMs like ChatGPT so good at their tasks. Unfortunately, real-world robotic data is scarce because there are limited real-world robot deployments. To address this data scarcity, some researchers have explored alternative approaches. Instead of relying on data collected from robots directly, they aim to teach robots by exposing them to videos of humans performing similar tasks. While this method holds promise, it remains uncertain how effective it can be. Humans learn through a combination of visual observation and physical experience. Simply watching a video of someone playing guitar, for instance, is insufficient for mastering the instrument. Practical application and hands-on experience are essential for true skill acquisition. This manifests both in terms of perception and cognition for and dexterity -- which we discuss in the next two sections.

Conclusion

While significant strides have been made in robotics, numerous challenges persist. Endowing robots with human-like intelligence, adaptability, and dexterity remains an ongoing pursuit.

We will dive into these issues in more details in future posts.

- Team Avsr AI