Why Smarter AI Doesn't Mean Smarter Robots (Yet)

The terms AI and robots are frequently conflated, but they represent distinct projects. Most headline-grabbing AI research operates within simulations or data centers, whereas robots contend with practical concerns like gravity, friction, battery life, and the inherent risks of objects falling. This article examines the current convergence of AI and robotics, the contributions of various AI techniques, and the primary obstacles standing in the way of more autonomous machines.

Robotics software typically consists of three layers: perception, planning, and control. AI has made significant strides in perception and planning, but control mechanisms are still predominantly classical. These control methods include PID loops, model predictive control, and inverse kinematics. The AI component operates on top of the control layer, rather than within it, which explains why even sophisticated vision systems may falter in executing tasks like picking up a cup due to the control layer's limitations.

Vision models represent the most mature AI contribution to robotics. They enable robots to detect objects, segment images, estimate depth from stereo cameras or LIDAR, and construct detailed 3D maps of their environment. Once the vision model produces a pose, the robot must then approach and grasp the object with the appropriate force, a separate challenge that is not addressed by the vision model.

Reinforcement learning (RL) is employed to train locomotion and manipulation policies. Instead of manually coding a walking gait, the robot simulates thousands or millions of movements, receiving rewards for maintaining balance and advancing forward. The policy emerges through this trial-and-error process. While RL has been used by companies like Boston Dynamics and Tesla for balance and recovery, the challenge of transferring these policies to real-world scenarios remains significant due to differences in friction, motor performance, and sensor accuracy in actual environments.

Vision-language-action models are emerging as planners rather than controllers in the realm of large language models (LLMs). These models take high-level instructions like "clear the table" and break them down into subtasks that are then processed by lower-level motion policies. For instance, Google's RT-2 model can translate images and text into robot actions, but the actual execution of these actions is still handled by the control stack.

Despite advancements in AI, robots are still constrained by power limitations, dexterity, cost, and safety considerations. A humanoid robot like Boston Dynamics' Optimus can operate for approximately 2-5 hours on a single charge, whereas human shifts typically last much longer. Human-like hands, with their 27 degrees of freedom and tactile senses, are something current robot grippers cannot replicate, which is why most warehouse robots manipulate boxes rather than individual items.

The cost of highly capable manipulation arms with force sensing capabilities is currently prohibitive, limiting the widespread adoption of humanoid robots in most workplaces. Additionally, safety is a paramount concern for robots interacting with humans, necessitating certified force limits and fail-safes that often add complexity to deployment compared to the AI algorithms themselves.

Moving forward, the next few years are likely to see robots become cheaper and more dependable at narrow tasks, rather than "smarter." The focus will shift from enhancing AI capabilities to improving the underlying hardware, including power efficiency, dexterity, manufacturing costs, and safety protocols.