Robotics today: hype or reality? | Navid Aghasadeghi | TEDxBoston

Key Concepts

• Moravex Paradox: The counterintuitive discovery that high-level reasoning is easy for robots, while sensorimotor skills (things humans find easy) are incredibly difficult.
• LLMs (Large Language Models): AI models like those powering ChatGPT, now being applied to robotics for generalized task understanding and planning.
• Teleoperation: A method of robot learning where humans remotely control robots to gather data on how tasks feel, enabling the robot to learn through experience.
• Evolution of Robot Software: Progressing from predefined movements to planning & perception, and now to AI-powered generalization.
• Embodied Data: The unique data robots need – sensory and physical feedback from interacting with the world – that doesn’t readily exist like the data used to train LLMs.

The Pivotal Moment in Robotics: From Promise to Potential

The speaker begins by acknowledging the long-standing anticipation of a robotics revolution, dating back to the 1920s, and questioning whether current advancements represent genuine progress or another “hype cycle.” Despite decades of research, the promised ubiquitous robots like Rosie from The Jetsons remain elusive. The speaker, with 15 years of experience at Rethink Robotics, Boston Dynamics, and in prosthetic control, attributes this delay to the inherent difficulty of robotics – a confluence of challenging problems in software, hardware, manufacturing, safety, and return on investment.

The Bottleneck: Software and the Moravex Paradox

The most significant bottleneck, the speaker argues, has been software. This is encapsulated by the Moravex Paradox, which states that tasks easy for humans (like walking or grasping) are incredibly difficult for robots, while tasks humans find hard (like chess) are relatively easy to program. He explains that building a chess engine capable of defeating a grandmaster was achieved before robots could reliably walk or grasp objects. This paradox highlights the difficulty of replicating human sensorimotor skills in machines. Numerous examples of robotic failures are cited as evidence of this historical challenge.

Three Phases of Robot Software Evolution

The speaker outlines a three-phase evolution of robot software:

1. Predefined Movements: Early robots operated on fixed paths (position A to B) with no environmental perception, limiting them to highly structured environments like automotive assembly lines. Any deviation of even a centimeter would cause failure.
2. Planning and Perception: This phase involved engineers designing complex algorithms for specific tasks, like a robot performing a backflip. Robots gained perception (e.g., Boston Dynamics robots identifying and jumping onto boxes) but remained limited to pre-programmed actions. Failures were still common.
3. AI-Powered Robots: The current era leverages the same AI technologies driving ChatGPT and image generation. These LLMs (Large Language Models) offer the crucial ability to generalize beyond specific tasks, representing a significant leap forward.

From Excel Sheets to LLMs: The Changing “Brain” of Robots

The speaker illustrates this evolution with a compelling analogy:

• Phase 1: Robots with an “Excel sheet for a brain” – executing pre-defined steps sequentially.
• Phase 2: Robots with a “chess engine for a brain” – solving complex equations for specific tasks.
• Phase 3: Robots with an “LLM for a brain” – capable of generalizing across tasks and domains.

He emphasizes that the same LLM powering conversational AI can now be applied to robotics, allowing a robot to understand a command like “clean the kitchen” and autonomously plan and execute the necessary actions. Similarly, advancements in image generation (DALL-E, Stable Diffusion) are being adapted to control robot movements, enabling complex continuous tasks like folding laundry – something previously impossible through hardcoding.

The Data Challenge and the Rise of Teleoperation

Despite these advancements, robotics isn’t “solved.” The primary remaining challenge is data. LLMs were trained on massive datasets of text and images readily available on the internet. However, robots require a different kind of data – embodied data – representing the feeling of performing a task, including joint positions, forces, and sensory input. This data is scarce.

To address this, researchers are focusing on teleoperation, where humans remotely control robots using joysticks, guiding them through tasks and allowing the robot to learn from the experience. This allows the robot to gather the necessary embodied data to improve its performance.

Approaching the End of the Moravex Paradox?

The speaker believes we are nearing the end of the Moravex Paradox, with robots becoming increasingly capable of performing tasks that were previously considered simple for humans. However, widespread adoption is not imminent. He draws a parallel to the Palm Pilot, suggesting robotics needs a comparable “iPhone moment” to achieve ubiquity.

Robotics as a Mirror to Humanity

The speaker concludes by emphasizing his motivation for working in robotics extends beyond simply building advanced machines. He is driven by a desire to understand humans – how we walk, manipulate objects, and interact with the world. He believes robotics can serve as a “mirror to humanity,” providing insights into ourselves as we strive to augment human capabilities with technology. His ultimate goal is to build technologies that enhance human flourishing.