Humanoid Robots
Figure 03 Cleans a Living Room Alone: Inside the Helix 02 Demo That Sparked a Fight With Elon Musk
Figure AI released a demo of its Figure 03 robot autonomously cleaning a living room using the Helix 02 AI model, performing tasks from surface wiping to TV remote operation. The clip triggered a public exchange between CEO Brett Adcock and Elon Musk over whether the demo was truly autonomous.
Figure AI released its most impressive demo yet on March 9: the Figure 03 humanoid robot, powered by the company's Helix 02 vision-language-action model, autonomously cleaning a cluttered living room. The robot sprayed and wiped a dirty table, scooped toys into bins, tossed pillows onto a couch, and turned off a television by pressing its remote control button, all without human intervention or teleoperation.
The demo sparked a public exchange between Figure CEO Brett Adcock and Elon Musk, who questioned whether the robot was truly autonomous. Adcock's reply was blunt: "Fully autonomous, not teleoperated." The clip racked up millions of views and reignited a core debate in the humanoid robotics industry: are these demos real, or just carefully staged theater?
AI-generated image
A home environment like those Figure 03 is designed to clean autonomously.
Key Stats
168 cm
Figure 03 Height
30 DOF
Degrees of Freedom
5 hrs
Battery Runtime
~$20K
Target Price
What the Helix 02 Demo Actually Showed
The two-minute video, posted to Figure's website and social media channels, showed Figure 03 navigating a real living room filled with toys, cushions, cleaning supplies, and furniture arranged in tight configurations. The robot performed a sequence of distinct tasks without pausing for instructions or switching between pre-programmed routines.
First, it picked up a spray bottle, applied cleaner to a table surface, then grabbed a towel hanging from a chair and wiped the table with sustained downward pressure. The towel handling stood out: the robot dynamically adjusted its grip as the fabric draped and shifted, a task that requires real-time tactile feedback rather than scripted motion.
Next came bimanual manipulation. The robot grabbed a storage bin with both hands, walked it to a table covered in toy blocks, and scooped them in. At one point, it tucked a container under one arm to free a hand for picking up smaller objects, a whole-body coordination behavior that suggests integrated planning across the upper and lower body.
The demo's most striking moments were the smaller details. Figure 03 tossed a throw pillow onto the couch with an accurate, relaxed motion. It picked up a TV remote, reoriented it in-hand, and pressed the power button with a single finger. It side-stepped through a narrow gap between a couch and coffee table while carrying objects. None of these actions are individually groundbreaking, but strung together in a continuous sequence without human input, they represent something new.
Key Insight
Helix 02 is a single end-to-end neural network that maps raw camera pixels directly to full-body motor commands. There is no separate module for navigation, grasping, or task planning. The system learns all of these behaviors from training data, and according to Figure, it improves simply by collecting more demonstrations.
Under the Hood: How Helix 02 Works
Helix is Figure's proprietary vision-language-action (VLA) model. The first version powered factory tasks on Figure 02 at BMW's Spartanburg plant. Helix 02, the version shown in the living room demo, processes visual input from six onboard cameras (including two palm-mounted cameras for close-range manipulation) and outputs motor commands for all 30 degrees of freedom in real time.
The architecture runs on dual low-power GPUs built into the robot's torso, with no cloud dependency for core perception and control loops. Speech interaction uses an onboard speech-to-speech pipeline. The system processes sensor data at twice the frame rate and one-quarter the latency of the Figure 02 system, with 60% wider field of view from upgraded camera modules.
AI-generated image
Precision actuator components and tactile sensor arrays representative of Figure 03's hand technology.
Figure 03's hands use 20 degrees of freedom across 10 fingers with tactile fingertip sensors capable of detecting forces as small as 3 grams. The hand actuators use frameless brushless DC motors with strain-wave and cycloidal gearing, a design choice that prioritizes torque density and backdrivability (the ability to comply with external forces rather than fighting them). This matters for household tasks where objects are fragile, surfaces vary, and grip force must adapt constantly.
The battery is a custom 2.3 kWh pack with 94% higher energy density than Figure 02's cells, UN38.3 certified for safety, with multi-layer protections against overcharge, thermal runaway, and impact damage. Active liquid cooling keeps the cells in optimal temperature range. Charging happens wirelessly through inductive coils in the feet at 2 kW, with 10 Gbps mmWave wireless data offload for uploading training data when docked.
| Spec | Figure 03 | Figure 02 | Tesla Optimus Gen 3 |
|---|---|---|---|
| Height | 168 cm | 170 cm | 173 cm |
| Weight | 60 kg | 70 kg | ~57 kg |
| Payload | 20 kg | 25 kg | 20 kg |
| Hand DOF | 20 (10 fingers) | 16 | 22 |
| Total DOF | 30 | 65+ | ~50 |
| Battery | 2.3 kWh / 5 hrs | ~1.2 kWh / 4 hrs | ~2 kWh / 4 hrs est. |
| Walk Speed | 1.2 m/s | 1.2 m/s | ~1.8 m/s |
| Charging | Wireless (2 kW) | Wired | Wired |
| Target Market | Home + Commercial | Commercial/Factory | Factory (internal) |
The Autonomy Question: Adcock vs. Musk
Hours after the demo dropped, Elon Musk posted on X: "Autonomous or remotely operated?" The question wasn't idle. Humanoid robotics has a credibility problem. Multiple companies have released flashy videos that turned out to be partially or fully teleoperated, with a hidden human puppeteering the robot's movements through a VR rig. The industry's history of staged demos means every new video faces immediate scrutiny.
Adcock responded directly, quoting Musk's post: "Fully autonomous." He followed up with a longer comment: "Fully autonomous, not teleoperated. This is 10x more impressive than another robot-MMA-stunt," a pointed reference to viral combat robot videos that prioritize entertainment over practical capability.
The exchange matters because it highlights a genuine technical divide. Tesla's Optimus program has focused on factory deployment at Tesla's own plants, where the environment is controlled, tasks are repetitive, and the robot can be monitored by Tesla employees. Figure's demo places the robot in an unstructured home environment performing varied, sequential tasks. The two companies are solving different problems, and the autonomy requirements for a living room are meaningfully harder than those for a production line.
The Teleop Trust Problem
• Why it matters: If demos are teleoperated, they prove hardware capability, not AI capability. A remotely operated robot cleaning a room tells you nothing about whether the AI can do it alone.
• How to verify: True autonomy demos should show the full sequence uncut, with no suspicious camera angles hiding a teleop station. Figure released the full clip, though independent third-party verification remains rare in the industry.
• The standard going forward: As humanoid robots enter homes, the bar for proving autonomy will need to rise. Consumers won't buy a robot that needs a remote operator.
Figure's Home Market Strategy
Figure 03 is the first humanoid robot explicitly designed for consumer homes from day one. While Figure 02 was built for factories (and still operates at BMW Spartanburg, where it has moved over 90,000 parts), Figure 03 introduces design choices that only make sense in a domestic context.
The exterior is covered in soft, multi-density foam with washable, removable textile coverings. The foam serves as anti-pinch protection, reducing the risk of injury during accidental contact. The textiles make the robot visually approachable rather than industrial, and they can be cleaned (an obvious need for a machine that handles dirty dishes and dusty surfaces). Speaker volume is doubled from Figure 02, and audio power is quadrupled, designed for conversational interaction across a noisy household.
AI-generated image
Figure's BotQ factory, where robots help build robots, targets 12,000 units per year initially.
The price target tells the real story. At scale, Figure is aiming for a retail price under $20,000, roughly the cost of a used car. Current unit economics are likely far above that number (early commercial humanoids from other companies have been quoted in the $100,000 to $200,000 range), but the BotQ factory is designed to close that gap. Initial capacity sits at 12,000 units per year, with a roadmap to 100,000 units by 2029.
The wireless charging system plays directly into the home thesis. Rather than plugging in a cable (awkward, trip hazard, requires precise docking), Figure 03 steps onto a floor pad and charges through inductive coils in its feet at 2 kW. The robot can park itself on the charger between tasks and resume work when needed, with no human intervention required.
🏠 Home-First Design Choices
Soft foam exterior, washable textile covers, wireless charging, enhanced audio for voice interaction, compact 168 cm frame for household navigation.
🏭 Factory Heritage
Figure 02's BMW deployment (90,000+ parts moved) proved the hardware and AI platform. Figure 03 adapts that proven base for unstructured home environments.
Who Else Is Targeting Homes?
Figure is not alone in eyeing the consumer market, but it is arguably the furthest along in demonstrating home-relevant capabilities with a production-designed robot.
• Tesla Optimus: Gen 3 production starts summer 2026 at Fremont. The focus remains internal factory use, with Musk's long-term vision of consumer sales at $20,000 to $30,000. No public home-environment demo exists yet. The new Gen 3 hands (22 DOF, 50 actuators) are impressive, but the AI stack is optimized for structured environments.
• 1X Technologies: The Norwegian company's Neo robot entered home testing in a few hundred households in 2025, with a world-model AI released in January 2026 that allows the robot to learn from visual observation. 1X targets a lower price point but has disclosed fewer technical details about home task performance.
• Sunday Robotics: Fresh off a $165 million Series B (March 2026) at unicorn valuation, Sunday is building purpose-designed home robots. The company is earlier-stage than Figure but has attracted serious capital for the home thesis.
• Honor: The smartphone maker debuted a small humanoid at MWC Barcelona in March 2026, positioned for shopping assistance and home companionship. The robot is less capable than Figure 03 but signals that consumer electronics companies see a market opening.
Key Insight
The home robot market is splitting into two camps: companies building general-purpose humanoids that happen to work at home (Figure, Tesla, 1X), and companies building home-specific robots from scratch (Sunday, Honor). The general-purpose approach is harder but potentially more valuable, since the same platform can serve commercial and consumer customers.
Frequently Asked Questions
When can I buy a Figure 03?
Figure has not announced a consumer release date. The robot is currently available to select commercial partners, with home alpha testing expected in 2026. At BotQ's initial capacity of 12,000 units per year, early availability will likely be limited. A broader consumer launch depends on scaling production toward the 100,000-unit annual target by 2029.
Is the $20,000 price realistic?
Not at current volumes. Early humanoid robots cost $100,000 to $200,000 per unit. The $20,000 target depends on high-volume manufacturing (tens of thousands of units annually), commodity component pricing, and the BotQ factory hitting its efficiency targets. Think of it like early EVs: the first Tesla Roadster cost $109,000; the Model 3 starts at $38,990 eight years later. The trajectory is plausible, but timelines are uncertain.
Was the Helix 02 demo really autonomous?
Figure CEO Brett Adcock stated publicly and repeatedly that the demo was fully autonomous with no teleoperation. The video appears to show continuous, uncut operation. Independent third-party verification has not been published. The robotics industry has a history of misleading demos, so healthy skepticism is reasonable, but Figure's track record with BMW deployment lends credibility to their claims.
What household tasks can Figure 03 actually do?
Based on the Helix 02 demo and Figure's published capabilities: surface cleaning (spray and wipe), picking up and sorting objects, carrying bins and containers, tossing soft items, operating simple controls (TV remotes, light switches), and navigating tight spaces between furniture. Tasks requiring fine motor skills (folding laundry, cooking) or heavy lifting (moving furniture) are not yet demonstrated at production quality.
The 12-Month Outlook
The Helix 02 living room demo is not proof that home robots are ready for mass adoption. It is proof that the gap between "factory-only" and "home-capable" is closing faster than most observers expected. Figure 03's combination of soft-body design, wireless charging, 5-hour runtime, and an end-to-end neural network that improves with data puts it in a different category from factory-floor prototypes that need controlled environments to function.
The milestones to watch over the next 12 months: BotQ production ramp (will they hit thousands of units?), home alpha testing results (how do real households respond?), and whether Helix 02's capabilities hold up outside of carefully selected demo environments. Tesla Optimus Gen 3 production starting this summer will also provide a direct comparison point, though Tesla's near-term focus remains internal factory deployment.
The Bottom Line: Figure AI just showed what a home robot needs to do. The question is no longer "can humanoid robots work in homes?" but "how fast can they get reliable and cheap enough for millions of them?"
The humanoid robot industry has spent years proving it can build machines that walk. The Helix 02 demo suggests the harder, more important problem, building machines that can actually do useful work in the places people live, is now within reach.