Figure 03 Home Robot: What the Helix Demo Proves

The short answer is that Figure 03 makes the demo more interesting, but not because it suddenly turns a humanoid into a consumer appliance. It matters because the new hardware claims line up with the weak spots that usually break home manipulation: occluded grasps, slipping objects, soft materials, household safety, docking, serviceability, and manufacturing cost.

That distinction is important for buyers. A robot that can tidy one staged room once is a demo. A robot with the sensors, hands, battery, charging, soft goods, and production plan to repeat that behavior in many homes is closer to a product roadmap. Figure is still on the roadmap side, not the shopping-cart side.

What is new compared with the older tidy benchmark?

The older living-room-tidy takeaway was about the task: clutter is a better test than dancing, backflips, or a single tabletop grasp. The newer question is about whether Figure 03's body fixes the specific reasons those demos usually fail.

Figure's official Figure 03 launch names three home-relevant changes that the previous benchmark could not evaluate directly:

• New hands and sensing: palm cameras, softer fingertips, and tactile sensors that Figure says can detect forces as small as three grams.
• Home usability work: soft textiles, multi-density foam near pinch points, washable/removable soft goods, better audio, wireless inductive charging, and battery safety work including UN38.3 certification.
• Manufacturing intent: a BotQ line that Figure says is initially capable of up to 12,000 humanoids per year, with a target of 100,000 robots over four years.

Those are not cosmetic claims. They are exactly the categories that separate a lab humanoid from a home robot. A household robot has to feel contact, see what its hand is doing when the head camera is blocked, survive contact with furniture, recharge without a human plugging it in, and be buildable at a cost that eventually reaches normal buyers.

The Figure 03 hardware claims that matter most

Figure's Helix 02 technical post says the system now combines a semantic reasoning layer, a fast visuomotor policy, and a learned whole-body controller. The headline numbers are useful: System 1 produces full-body joint targets at 200 Hz, System 0 runs at 1 kHz for balance and contact, and Figure says System 0 was trained on more than 1,000 hours of human motion data across more than 200,000 parallel simulation environments.

That architecture is interesting, but home buyers should anchor on the hardware claims because hardware is where household work gets messy.

The most important thread is redundancy. A home robot needs multiple ways to know what is happening: vision from the head, close-range hand vision, tactile feedback, joint state, force sensing, and language-level task context. If one of those channels fails, the robot should slow down, ask, retry, or stop. That is a much higher bar than completing a clean take for a launch video.

What the living-room demo still proves

Figure's living-room clip is still worth taking seriously. The company says Helix 02 learned the new behavior by adding data, with no new algorithms or special-case engineering. In the demo, the robot uses a spray bottle, wipes a surface, manages a towel, holds a bin with two hands, scoops blocks into it, tucks a container under one arm, tosses a pillow, reorients a TV remote, presses a button, stows tools while walking, and side-steps through a narrow path.

Those actions are not random. They are the right primitives for a home humanoid:

• loco-manipulation, because the robot moves while carrying or using objects;
• bimanual handling, because many chores need both hands;
• soft-object interaction, because towels, pillows, clothing, and blankets do not behave like blocks;
• in-hand adjustment, because a remote, bottle, cup, or toy is rarely already aligned correctly;
• temporary object storage, because real chores often require freeing one hand in the middle of the task;
• tight-space navigation, because homes are not open factory aisles.

This is stronger evidence than a humanoid walking on stage. It is also not proof that Figure 03 can tidy your living room next year. The demo does not publish success rates, number of attempts, runtime, intervention count, child or pet interruptions, privacy settings, cleaning-material limits, fragile-object policies, or what happens when the room is rearranged.

The Figure 02 factory baseline helps, but only so much

Figure has one advantage many home-robot claims do not: an industrial deployment history. In the ui44 database, Figure 02 is marked as a retired industrial humanoid with a 168 cm body, 70 kg weight, 20 kg payload, 16-DOF hands, and a BMW Spartanburg deployment where Figure says the robot contributed to more than 30,000 cars across 1,250+ runtime hours.

That matters because factory runtime is better than a pure lab reel. It suggests Figure can ship hardware into a controlled worksite, gather data, retire a platform, and iterate to a new generation.

But the home is less forgiving than a factory in different ways. A plant has trained staff, marked work cells, emergency procedures, known tasks, and a business case for supervision. A home has pets, glassware, loose cables, children, spilled food, awkward furniture, and objects the robot should not move at all. The home problem is not just manipulation. It is permission, judgment, and support.

How Figure 03 compares with other home-manipulation paths

The ui44 database is useful here because Figure is not the only route toward home manipulation. A walking humanoid is one architecture. A wheeled mobile manipulator, a lighter home humanoid, or a compact developer platform may be more realistic for early buyers.

Stretch 4 is the useful reality check. It does not need legs to reach into a home environment. It has a real list price, a self-charging dock, open ROS 2 and Python tooling, and a clearer research/assistive pathway. If your question is "what can be tested in homes now?", a wheeled mobile manipulator may be more credible than a bipedal humanoid with a stronger video.

1X NEO is the other comparison because it is explicitly home-focused. In ui44, NEO is much lighter than Figure 03 at 30 kg versus 61 kg, with a soft body, tactile skin, and a public $20,000 early-adopter price. That does not make NEO solved, but it shows the trade-off: Figure's evidence is technically deeper around full-body autonomy, while 1X has clearer consumer positioning.

What evidence is still missing before this is a home product?

A publishable home-robot claim needs more than a chore list. For Figure 03, the next useful evidence would be:

1. Repeat trials across different homes. The same task in ten ordinary rooms would mean more than one polished environment.
2. Success and failure numbers. Buyers need completion rate, drop rate, collision rate, intervention rate, and average task time.
3. Human-presence tests. The robot should show what happens when a person, child, or pet enters the scene.
4. Permission controls. Cleaning, moving objects, pressing buttons, opening storage, and handling medicine need user-set boundaries.
5. Privacy controls. Palm cameras, head cameras, audio, and data offload are powerful; they need clear retention and opt-out rules.
6. Service terms. Hands, tactile fingertips, textiles, batteries, and docks will wear. The buyer question is who repairs them and at what cost.
7. Launch terms. No public consumer price, support plan, delivery region, or in-home warranty means this is not yet a buying decision.

That is not a criticism of Figure specifically. It is the right bar for every home-humanoid company. The closer a robot gets to useful chores, the more it needs boring product answers.

Should home-robot buyers care now?

Yes, but as evidence, not as a purchase signal.

Figure 03 and Helix 02 deserve attention because the story is now more coherent than a typical humanoid hype cycle. The software demonstrates long-horizon, whole-body manipulation. The hardware adds tactile fingertips, palm cameras, softer materials, inductive charging, and home-safety design work. The company also has factory deployment experience with Figure 02 and a stated manufacturing plan through BotQ.

The caveat is just as clear: Figure 03 is not listed in the ui44 database as a consumer-available robot, has no public price, and has no published consumer support model. A buyer should not treat the living-room demo as proof that a robot can safely clean an arbitrary home unattended.

The practical takeaway is this: Figure has moved the conversation from "can a humanoid perform a neat trick?" to "can a humanoid hardware-and-AI stack become a repeatable home product?" That is real progress. It is also still a question, not an answer.