Figure AI has shared footage of its Figure 03 humanoid robot climbing a ladder fully autonomously — a task that demands the kind of whole-body coordination, balance, and real-time spatial awareness that has long been a frontier problem in robotics.
The video, posted on X by Figure founder Brett Adcock, shows the robot ascending a ladder without visible human assistance or remote control. According to Interesting Engineering, the demonstration showcases what the company describes as another milestone in robotic mobility and balance.
🔍 THE BOTTOM LINE
Ladder-climbing is one of the hardest things a humanoid robot can do. It requires simultaneous coordination of arms and legs, continuous balance adjustments, and accurate perception of the environment — all while supporting the robot’s full weight on narrow rungs. Figure 03 pulling this off autonomously, with no safety tethers visible, signals that whole-body control is moving from the lab into the realm of practical capability.
Why Ladder-Climbing Matters
Climbing a ladder is not just a parlor trick. It is a benchmark that exposes every weakness in a humanoid robot’s control system. The robot must:
- See and interpret the ladder — rung spacing, angle, surface texture, surrounding obstacles
- Plan foot and hand placement — choosing where to grip, in what sequence, with what force
- Maintain balance throughout — shifting centre of mass with every step, compensating for the dynamic load of moving limbs
- Recover from slips — adjusting in real time if a foot misses a rung or a handhold shifts
A robot that can walk on flat ground is impressive. A robot that can climb a ladder is operating in a different regime of difficulty — one that demands the same combination of perception, planning, and physical execution that humans use unconsciously every day.
The Helix S0 Upgrade: Vision Meets Motion
The ladder climb is powered by Figure’s recently upgraded Helix System 0 (S0) AI model, which now combines real-time visual perception with whole-body motion control.
Previously, the S0 system relied only on proprioception — the robot’s awareness of its own joint positions, body movements, and balance. That approach works for basic locomotion on flat surfaces but breaks down when the robot encounters stairs, ladders, or uneven terrain where it needs to actively see and interpret its surroundings.
The updated model processes RGB images from onboard stereo cameras to build a three-dimensional representation of the environment. The robot simultaneously sees the terrain ahead and monitors its own body position, enabling more precise foot placement and smoother movement across challenging surfaces.
Figure says the model was trained end-to-end using reinforcement learning in simulation, with behaviours that transfer directly to real-world robots without additional calibration. This addresses the long-standing “sim-to-real” gap — the challenge of getting policies learned in simulation to work reliably in the messy, unpredictable real world.
Production Scale-Up
The ladder demonstration comes alongside news that Figure has increased production of the Figure 03 from one unit per day to one per hour, delivering over 350 robots. The company is scaling manufacturing while simultaneously advancing the AI model that powers them — a combination that matters because each new robot in the field generates data that feeds back into model improvement.
Adcock separately argued that “wheeled robots are an utter dead end,” reiterating his belief that legged humanoids are better suited for environments designed for humans — factories, warehouses, offices, and homes where stairs, ladders, and uneven surfaces are the norm.
The Competitive Context
Figure is not alone in pushing the boundaries of humanoid mobility. The sector is crowded: Tesla’s Optimus is in production with over 1,000 units deployed, BYD is unveiling its first humanoid this month, and NVIDIA has released an open foundation model for humanoid robots. Chinese companies like Unitree and UBTech are racing to mass-produce humanoids at lower price points.
What sets the Figure 03 demonstration apart is the specific claim of autonomous ladder-climbing — a task that sits in the gap between what most humanoids can do (walk, carry objects, perform repetitive assembly tasks) and what they aspire to do (navigate any environment a human can). The broader trend across humanoid robotics is moving from single-task demonstrations toward general-purpose autonomy, and each new capability milestone narrows that gap.
What Has Not Been Verified
The demonstration has not been independently verified. Figure has not disclosed technical details about the underlying system — sensor configurations, model architecture, or failure rates. The video shows a controlled environment, and it is unclear how the robot would perform on different ladder types, in poor lighting, or under load.
AI researchers described the demonstration as “good progress” in humanoid robotics, which is a measured assessment. Progress in robotics is incremental, and each milestone builds on the last. The gap between climbing a ladder in a demo and climbing a ladder reliably, thousands of times, in a working factory, remains substantial.
❓ FAQ
Is the Figure 03 robot available commercially? Figure is producing the Figure 03 at a rate of one unit per hour and has delivered over 350 robots. The company has not publicly disclosed pricing or general availability for consumers. Its focus appears to be industrial and commercial deployments.
How is this different from other humanoid robot demonstrations? Most humanoid demonstrations focus on walking, carrying objects, or performing assembly tasks. Ladder-climbing requires a fundamentally different control challenge — the robot must support its full weight on narrow rungs while coordinating four limbs in sequence. The combination of visual perception and whole-body motion control is the technical advance here.
What is the “sim-to-real” problem? Training robots in simulation is cheap and fast, but simulations never perfectly match reality — physics, friction, sensor noise, and lighting all differ. The sim-to-real gap is the challenge of getting AI policies learned in simulation to work reliably on physical robots. Figure claims its ladder-climbing behaviour transfers from simulation to reality without additional fine-tuning, which, if verified, would be notable.
Could this technology be used in New Zealand? New Zealand’s manufacturing and logistics sectors face ongoing labour shortages, particularly in physically demanding roles. Humanoid robots that can navigate human environments — including ladders and stairs — could eventually be deployed in warehouses, factories, and construction sites. The technology is not ready for widespread commercial deployment yet, but the trajectory matters for any country reliant on physical industries.
🔍 THE BOTTOM LINE
A robot climbing a ladder on its own is a genuine technical milestone — not because ladders are common in factories, but because the capability signals that humanoid robots are approaching the kind of general-purpose mobility that would make them useful in any environment a human worker can navigate. The gap between demo and deployment remains wide, but it is narrowing.
📰 Sources
- Interesting Engineering — Figure 03 humanoid climbs ladder autonomously
- Brett Adcock (X) — Figure 03 ladder demo video
- Machine Dawn — Today in Robotics, 4 August 2026
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
