Physical AI
Figure AI's Robot Ran 67 Hours Straight Without Help
Figure AI announced that its Figure 03 robot completed 67 consecutive hours of unsupervised autonomous operation with just one error, powered by the new Helix 2 neural network. The milestone, tested across home and commercial environments, represents the longest publicly documented stretch of continuous humanoid robot autonomy.
Figure AI announced on March 15 that its robots completed 67 consecutive hours of unsupervised autonomous operation, logging just one error across the entire run. The milestone, powered by the company's new Helix 2 neural network model, marks the longest publicly documented stretch of continuous, hands-off humanoid robot autonomy to date.
The test took place across home and commercial environments using the Figure 03 platform. Rather than relying on hundreds of thousands of lines of traditional code, Helix 2 uses a single neural network for full-body control, enabling the robot to learn and generalize skills through data rather than hand-coded instructions. CEO Brett Adcock put it simply: "If you can teleoperate the robot to do a task, you can train the neural net to learn it."
AI-generated image
Neural network data flow visualization. Source: biped.news
Key Stats
67 hrs
Continuous Autonomy
1 error
Over Entire Run
$39B
Company Valuation
12,000
Annual BotQ Capacity
What Helix 2 Actually Does
Helix 2 is Figure AI's second-generation vision-language-action (VLA) model. It replaced the company's earlier approach, which layered separate systems for perception, planning, and motor control. Now a single neural network handles everything from camera input to joint-level actuation. The result is smoother, more fluid movement and faster reaction times compared to Helix 1, which Figure deployed internally in late 2025.
The practical effect is significant. Where previous systems required engineers to anticipate edge cases and write specific handling code, Helix 2 generalizes from demonstration data. A human operator teleoperates the robot through a task once or a few times, and the neural net learns to replicate and adapt the behavior across varied conditions. This includes tasks like loading dishwashers, sorting laundry, clearing tables, and organizing shelves.
🧠 Single Neural Network Control
One model handles vision, language understanding, and motor actions. No separate modules for perception, planning, and execution.
📊 Learn From Demonstration
Teleoperate a task once, and the network generalizes the skill. No manual coding required per-task.
Figure claims eight new autonomous skills were demonstrated in the days leading up to the 67-hour test, including what the company called "world-leading cleaning" with coordinated multi-tool use. The robot picked up objects, operated appliances, and navigated between rooms without human intervention for nearly three full days straight.
Under the Hood: Figure 03's Hardware Stack
The 67-hour run used Figure 03, the company's third-generation humanoid announced in October 2025. At 168 cm tall and 60 kg, it is 9% lighter than Figure 02. The weight reduction comes from a combination of redesigned actuators (frameless brushless DC motors with strain-wave gearing, running at 2x the speed of Figure 02's) and a shift from CNC-machined parts to die-cast and injection-molded components built for volume production.
The hands are a standout feature. Each has five fingers with 20 total degrees of freedom across both hands, plus palm-mounted cameras with wide field of view and low latency. Tactile sensors detect forces as small as 3 grams, enough to sense the moment an object starts to slip. The vision system runs at double the frame rate and one-quarter the latency of Figure 02's cameras, with 60% more field of view per sensor.
AI-generated image
High-tech manufacturing environment representative of Figure AI's BotQ facility. Source: biped.news
Onboard compute runs dual embedded GPUs processing the Helix VLA model locally, with no cloud dependency for core manipulation. A 10 Gbps mmWave wireless link handles data offload for fleet learning. Battery life sits at roughly 5 hours per charge, with wireless inductive charging through foot-mounted coils at 2 kW.
| Spec | Figure 03 | Tesla Optimus Gen 3 | Boston Dynamics Atlas |
|---|---|---|---|
| Height | 1.68 m | 1.73 m | 1.88 m |
| Weight | 60 kg | ~57 kg | ~89 kg |
| Hand DOF | 20 | 22 | Not disclosed |
| Total DOF | 30 | ~28 | ~28 |
| Battery Life | ~5 hrs | ~24 hrs (claimed) | Not disclosed |
| AI Model | Helix 2 VLA | FSD-v15 | Proprietary |
| Longest Autonomous Run | 67 hrs (1 error) | Not publicly tested | Not publicly tested |
| Est. Unit Price | ~$130,000 | Target <$20,000 | ~$200,000 |
| Production Status | BotQ ramping | Fremont production | 30K/yr committed |
The Autonomy Race: Who Else Is Competing
Figure's 67-hour benchmark lands in a week already packed with humanoid robotics news. The competitive picture is getting sharper, and each major player is staking out a different strategic position.
• Tesla Optimus Gen 3: Debuted at AWE 2026 in Shanghai on March 13 with 22-DOF hands and mass production underway at Fremont. Tesla is targeting 50,000 to 100,000 units per year initially, with a long-term price target below $20,000. The approach prioritizes scale and cost over autonomy duration.
• Boston Dynamics Atlas: The electric version entered full production after winning "Best Robot" at CES 2026 in January. All 2026 units (up to 30,000) are committed to Hyundai factories and Google DeepMind pilot programs. At roughly $200,000 per unit, Atlas sits at the premium end of the market.
• Agility Robotics Digit: Already deployed at Toyota Motor Manufacturing Canada for RAV4 production line logistics, with additional rollouts at Mercado Libre's Texas facility. Digit is the most commercially deployed humanoid right now, backed by $641 million in funding from Amazon and SoftBank.
• Sunday Robotics: Raised $165 million on March 12 at a $1.15 billion valuation, targeting household robots with its "Memo" platform. Still pre-deployment, but the funding signals investor appetite for home-focused humanoids.
Key Insight
Figure is the only company publicly benchmarking multi-day autonomous operation. Tesla and Boston Dynamics have demonstrated impressive hardware, but neither has released comparable endurance data. This matters because the gap between a 10-minute demo and a 67-hour deployment is where real commercial viability lives.
What 67 Hours of Autonomy Means for the Industry
Duration metrics have become a proxy for readiness. A robot that operates for an hour under supervision is a research project. One that runs for nearly three days with minimal errors starts to look like a viable product. The 67-hour test, if the numbers hold up under independent verification, pushes Figure past the threshold where early customers can realistically plan around robotic labor for specific tasks.
AI-generated image
Smart home environment with integrated sensing, representing Figure's target deployment context. Source: biped.news
The home deployment angle is particularly notable. Figure plans to begin alpha testing Figure 03 in actual homes during 2026, followed by scaled pilot programs in 2027 and 2028, and mass production by 2028 to 2029. This puts the company on roughly the same timeline as Tesla's home deployment ambitions, but with demonstrated long-duration autonomy data that Tesla has yet to publish.
Manufacturing scale is the other half of the equation. Figure's BotQ factory in San Jose can produce 12,000 units per year on its first-generation line, with plans to scale to 100,000 units over four years. That is smaller than Tesla's stated 50,000 to 100,000 annual target, but Figure is building robots that cost roughly six times more per unit. The revenue math could be comparable even at lower volumes.
The Self-Replicating Angle
Figure is using its own robots on the BotQ production line, essentially building future robots with current ones. Investor Peter Diamandis noted: "Every improvement makes it better at building the next generation." This bootstrapping approach, if it scales, could compress manufacturing timelines significantly.
Frequently Asked Questions
What does "67 hours of autonomous operation" actually mean?
The robot performed household and commercial tasks for 67 consecutive hours without human intervention, guidance, or remote control. It navigated between rooms, manipulated objects, operated appliances, and handled cleaning tasks using only its onboard AI. One error occurred during the entire run, though Figure has not disclosed the specific nature of that error.
How does Helix 2 differ from traditional robot control software?
Traditional approaches use separate software modules for perception (seeing), planning (deciding), and control (moving). Each module has thousands of lines of hand-written code. Helix 2 replaces this stack with a single neural network that maps camera input directly to joint movements. Skills are learned from human demonstrations rather than programmed rule by rule.
When can consumers buy a Figure robot for home use?
Not soon. Figure plans alpha testing in homes during 2026 with a small number of units, scaled pilot programs through 2027 and 2028, and mass production targeting 2028 to 2029. At an estimated $130,000 per unit today, early access will likely be limited to pilot program participants and enterprise customers. Pricing is expected to drop as manufacturing scales.
How does Figure's valuation compare to competitors?
Figure AI was valued at $39 billion after its September 2025 Series C round of over $1 billion. For context, Agility Robotics has raised $641 million total, and Sunday Robotics just hit $1.15 billion in valuation. Tesla's robotics division is not separately valued, but analysts peg Optimus as potentially worth hundreds of billions if mass production targets are met.
What's Coming Next
The next six months will determine whether the 67-hour benchmark translates into commercial traction. Figure expects to place units in real homes for alpha testing before the end of 2026, which will generate the first uncontrolled-environment data at scale. Production at BotQ needs to prove it can hit the 12,000-unit annual run rate. And competitors will respond with their own endurance claims.
Watch for three specific signals: whether Figure releases third-party verification of the autonomy test, how quickly home alpha participants report back on real-world reliability, and whether Tesla or Boston Dynamics publish comparable duration benchmarks. The company that cracks multi-day unsupervised operation in unstructured environments holds the strongest card in a market projected to reach trillions of dollars by the early 2030s.
The Bottom Line: Figure AI's 67-hour autonomy run is the first public proof that a humanoid robot can work for days, not minutes, without help. If BotQ production scales and home alpha tests confirm the data, Figure holds a measurable lead in the metric that matters most: how long a robot can be useful before a human needs to step in.