Do robots really need legs?

In this talk

1. 1:30

   Choose a robot form for the job

   Pierce says legged robots will have some uses, but many current factory and warehouse tasks take place on flat floors. He expects mobile robots to tackle structured, repetitive work first, while humanoids that can do anything at home are likely much farther away. He compares humanoids to Ferraris and simpler robots such as Roombas to Honda Civics.

2. 7:14

   Prioritize manipulation and practical tools

   Pierce says getting robots into simpler workplaces can generate data while teams work on the difficult “last 30-centimeter” problem: grasping objects and using tools. Human-like hands are mechanically complex and can cost around $15,000 each, so his company is exploring interchangeable tools such as vacuum grippers, tweezers, and parallel grippers. He suggests purpose-built kitchen tools that snap onto a robot could be simpler than making a hand work with every utensil.

3. 16:21

   Latency and compute limit robot speed

   Pierce says robot motors can move quickly, but cameras, planning models, and safety checks add latency between seeing a scene and acting. Processing feeds from several HD cameras and coordinating high-level decisions with faster controllers takes substantial onboard computing. He says some well-funded companies use H100 or H200 chips on robots, and hopes that capability could cost about $500 in five years.

4. 20:04

   AI models make pilots easier to deploy

   Pierce argues that recent progress builds on large language models and vision systems developed by companies including Meta, Google, and Nvidia. His team can use AI to read warehouse screens that human operators already use, avoiding custom integration with warehouse software. He says this can get a pilot running in two days, compared with earlier projects that could take months before a pilot began.

5. 24:01

   Manufacturing and rules shape what comes next

   Pierce sees the United States and China as the leading contenders, combining strong American AI with China's manufacturing capacity. He says his company has struggled to find local suppliers for components and tooling, noting that injection-molding tools are often made in China. Looking ahead, he expects rapid lab progress but says regulation and legal concerns, including what happens when an emergency stop makes a humanoid fall, may slow deployment.