Figure 01 and the Path Toward More Capable Robots

I am sick, so I cloned my voice rather than miss this week’s AI news. First is a Figure AI demonstration in which Figure 01 uses OpenAI technology to converse while interacting with objects on a table. This is a demonstration of specific capabilities, not evidence that the robot has achieved AGI.

In the Figure clip, a person asks what the robot sees. Figure 01 identifies a red apple on a plate, a drying rack with dishes, and the person nearby. Asked for something to eat, it hands over the apple. While handling trash, it explains that the apple was the only edible item it could provide from the table.

The person then asks where the dishes should go. Figure 01 says the plate and cup should go into the drying rack and proceeds to move them there. Asked to assess the task afterward, the robot says the apple has a new owner, the trash is gone, and the tableware is where it belongs. This is a presented tabletop demonstration; the clip alone does not establish how reliably it performs outside that setting.

Next I show a Unitree humanoid moving quickly; I describe it as a humanoid speed record, although the clip does not independently establish that claim. A Boston Dynamics robot follows, moving heavy objects with precision. To me, Boston Dynamics still looks ahead in physical ability.

Another clip shows a robot navigating very tight spaces and uneven rubble. I can imagine such a machine being useful for a cave rescue or other confined work, although the video is a demonstration rather than proof of a completed rescue. Another demonstration shows a person using an Apple Vision Pro headset as a robot controller.

A separate clip visits a robotics startup in a Palo Alto garage. An interviewer asks about its roughly two-month-old robot, and a team member says they are proud it can stand without falling over. The team is trying to iterate on hardware quickly with 3D printers: the machine shown is version three, and they say a fourth version could be made within days of finishing its design.

Asked when a robot like this might clean, do laundry, or wash dishes at home, the team member says the desire to have one personally motivated the project. For now, the robot is a test bed for implementing recent research, including the UMI paper; the timeline for doing so is still to be determined. I see this as part of a broader trend toward cheaper robot development.

The next clip is not a robot: it shows an all-electric passenger-carrying aircraft in flight. After that I show footage associated with NIO, the Chinese electric-vehicle company, including robotic work in a factory. A short promotional clip does not by itself establish the level of autonomy across a factory.

Finally, I play part of an interview between Lex Fridman and Yann LeCun. Asked whether millions of humanoid robots will be walking around soon, LeCun says not soon, but expects the next decade to be interesting for robotics. In his view, a central problem is getting systems to understand how the world works and plan actions.

LeCun says specialized tasks can be tackled with carefully planned, handcrafted dynamical models, as in classical robotics, alongside innovation and perception. But he contrasts those achievements with building a domestic robot. He also argues that fully autonomous level-five driving remains difficult, especially if a system must learn with the limited driving experience a teenager might need.

LeCun believes significant progress in robotics depends on world models that systems can learn to understand how the world works. Hardware developers are betting that AI will progress enough to make useful products possible. Fridman suggests that before a very strong world model arrives, people may still find uses for a clumsy, imperfect robot.

Fridman raises factories as one possible setting, while LeCun cautions that safety requirements make factory deployment very difficult. They also discuss homes, where clearing a table, washing dishes, cleaning, and cooking are appealing targets. LeCun emphasizes that these apparently ordinary tasks are sophisticated and complicated.

Fridman asks about navigation in uncertain spaces. LeCun says basic navigation works to a degree, but navigating in a way humans find compelling is another matter. At Meta’s embodied AI group, researchers have used commercial robot dogs in demonstrations: a robot can navigate to a fridge, open it, and potentially retrieve a recognized can.

LeCun distinguishes these abilities from a general robot sophisticated enough to clear a dinner table. Fridman finds the prospect of robots in homes exciting because people could interact with AI physically and explore the psychological and philosophical dimensions of those relationships. LeCun says his team has worked on self-supervised learning from video for around a decade, with significant progress in the preceding two or three years.