Figure 03 Does Chores — How It Recovers When Stuck Is What Matters
Community Discussion · Humanoid Robots

Figure 03 Does Chores — How It Recovers When Stuck Is What Matters

YanshiYanshi2d ago2026/11/17 96 views

Spent the weekend messing with a Figure 03 borrowed from a client, and stepped in plenty of pits. Only two days, the venue a forty-something-square-meter showroom set up like a living room, with sofa, coffee table, rug, and dining table all in place. The manufacturer's engineer stayed for two hours, gave the rundown, and left — the rest we ran ourselves.

Overall, its action chain is complete. But complete and stable are two different things, let alone hands-off.

First impression is the form factor. About 9% lighter than the previous Figure 02, and a size smaller — it fits through ordinary door frames and kitchen aisles without turning sideways. Key parts use multi-density materials; the palms and the outside of the fingers are a soft layer that feels like thick silicone. The benefit is concrete: it doesn't leave marks when it bumps a coffee table leg, and standing next to a person it's less intimidating. The first thing I did was find the e-stop — the button isn't hard to locate, and slapping it cuts power immediately. That matters to me.

On degrees of freedom I got the simplified spec sheet and didn't dig in. From what I saw, wrist and forearm rotation is sufficient — it handles glass cups, sponges, folded clothes. When grabbing a cup you can tell it's modulating force, not clamping rigidly.

Noise is lower than expected. Standing a meter away it's mostly the low-frequency hum of joint servos, not as loud as those old hydraulic setups.

What really interested me was failure recovery. The sponge slipped out of its hand once; it looked down for about two seconds, picked it up again, and carried on. That matters far more than a single success — dropping things is normal in a household scenario, and being able to clean up after itself is what makes it usable.

Following what the engineer taught me, I framed the target object and picked the task on the tablet, and it did the rest on its own. I can't remember exactly what the interface looks like, but roughly camera feeds on the left and a row of tasks on the right. The flow isn't hard to pick up — half an hour and you can operate it yourself.

There were plenty of places it got stuck.

Floor. At rug edges and thresholds it hesitates and stops waiting for a command. I had to manually give it a pass command before it continued. Not fatal, but it means the floor at home can't be too fragmented.

Lighting. In the afternoon when sunlight slanted in onto the dining table side, recognition slowed noticeably and it took two retries to get it right. The vendor's demo videos are basically shot under even lighting; a real home isn't like that.

Speed. For the dishwashing routine I timed it — an order of magnitude slower than a human. Slow isn't wrong per se; the problem is you have to watch it the whole time, and that time cost has to be counted.

The hardest thing to pin down is the ratio of autonomy to teleoperation. The engineer said it was running autonomously on site, but didn't give an intervention count. A while back a report questioned its demo video, with the line "beneath this video that set the internet ablaze is a shadow war over technical authenticity and the industry's future." I don't think it's faked, but an action sequence that can run through once and a system that can run a hundred times in a row are separated by a lot of engineering work. On battery life they only said it depends on the task load and didn't give specific hours, so I can't verify that either. Also, I only just started reading up on vision-based foundation policies these two days, so I don't dare judge how far it's gotten on the end-to-end path.

Whether it's worth deploying depends on the scenario. If you have a clear list of repetitive tasks, flat floors, and can accept a technician watching nearby, this machine is usable now — showrooms, model homes, early pilots in high-end residences all count. It made Time's annual list, and the reception value from that exposure is real. If you're expecting it to independently watch the house and mind the kids — forget it. If your home has lots of rugs and thresholds and complex lighting, hold off. And if you're budget-sensitive, don't even bring it up — at this stage the price and delivery pace don't suit ordinary households.

Figure 03 right now is a prototype that can run a process through end to end. If you really want it to work independently, you'll have to wait a while longer.

2 replies

?
Ctrl + Enter to reply
Production Line Old Li

Being able to pick the sponge back up by itself after slipping is nice, but at the carpet edge it still needs a human to give commands — on our production line, how long a changeover can it handle?

Demo Still Far

"Failure recovery" matters more than one-off success, but the carpet edge still needs a manual pass command — that's a whole engineering effort away from hands-off.