Can Humanoid Robots Climb Stairs?

Figure AI's new System 0 claim makes the question worth revisiting. In its latest Figure 03 production update, Figure says its humanoid can now use camera perception inside its whole-body controller to traverse stairs, ramps, and uneven terrain zero-shot from simulation, with no real-world fine-tuning and no operator-in-the-loop adjustment. That is a serious mobility claim. It is not, by itself, proof that you should put a humanoid on your staircase.

The honest answer is: some humanoid robots can climb stairs in demos or research settings, but no broadly available home humanoid has proven safe, repeatable, consumer-ready stair autonomy yet. Stairs matter because they are not just a locomotion trick. They force perception, balance, recovery, payload, safety policy, and support infrastructure to work together in a place where a mistake can damage a robot, a home, or a person.

Can humanoid robots climb stairs today?

Yes, but with a giant asterisk. Stair climbing is not new to humanoid robotics. Honda's ASIMO could climb stairs years ago. PAL Robotics' REEM-C, a 165 cm, 80 kg research humanoid, is listed in the ui44 database with stable walking, stair climbing, chair sitting, ROS support, 68 degrees of freedom, and a 3-hour walking battery figure. Unitree's full-size H1 is recorded with stair-climbing and terrain adaptation among its capabilities.

That history is useful because it prevents hype. A humanoid stepping up a known staircase does not automatically mean it can roam a normal home. The harder question is whether it can notice an unfamiliar staircase, estimate the height and depth of each step, keep its body stable, carry an object, pause when a pet crosses its path, recover from a foot placement error, and explain or log what happened if it refuses to continue.

Figure's System 0 update is interesting because it points at exactly that transition. Figure says S0 previously reasoned only about the robot's own body: joint state, base motion, and proprioception. It could walk on flat ground, but stairs and ramps required hand-tuned mode switches and operator intervention. The new version feeds RGB images from the onboard head cameras through a stereo model to build a 3D scene representation, then conditions the whole-body policy on that scene plus proprioception. In plain language: the robot is not just feeling its body; it is looking at the terrain before it steps.

That is the right direction for homes. A household staircase is not a factory lane. It may have carpet, glossy wood, shadows, a turn, a baby gate, shoes on the landing, poor lighting, or a person coming the other way. A controller that cannot perceive the world has to rely on scripted assumptions. A controller that can perceive still has to prove it can act safely when those assumptions break.

Why Figure System 0 is more than a stair demo

The most important part of Figure's announcement is not simply "the robot climbed stairs." It is the claimed architecture: perception in, whole-body control out, trained in simulation across thousands of randomized terrains, then deployed on hardware zero-shot.

For buyers, that matters for three reasons.

First, it hints at generalization. If a robot needs a special hand-coded routine for every staircase, it will struggle in real homes. If the same policy can use 3D scene understanding for stairs, ramps, thresholds, and uneven flooring, the robot starts to look less like a demo machine and more like a mobile helper. Figure is explicit that stairs are the first use case, not the whole point.

Second, it connects to manufacturing scale. Figure says BotQ has delivered over 350 Figure 03 units, increased production from one robot per day to one per hour, produced more than 9,000 actuators, reached over 80% end-of-line first-pass yield, and subjects each robot to more than 80 functional verification tests. Those numbers matter because mobility failures are often long-tail failures. You do not learn all of them from one beautiful video. You learn them from many robots, many hours, logs, repairs, and repeated failures.

Third, it shows why the home story still needs restraint. Figure 03 is listed in the ui44 database as an active humanoid with a 173 cm height, 61 kg weight, roughly 5-hour battery life, 4.3 km/h maximum speed, stereo vision, depth cameras, force sensors, tactile arrays, Helix VLA, and a 20 kg payload. It is also not available for consumer purchase and has no public price. The same facts that make Figure 03 exciting also make it a fleet-development platform first, not a normal home product.

That is why stair climbing should be read as a benchmark, not a purchase signal. A company showing perception-conditioned stair traversal is telling us something about the maturity of its control stack. It is not yet telling us warranty terms, liability policy, child safety behavior, remote-support rules, or whether the robot is allowed to use your staircase while carrying laundry.

What does the ui44 database say about stair-ready robots?

The database shows a messy split: many robots have impressive mobility, but few connect that mobility to a supported home use case.

Unitree H1 is the cleanest public humanoid comparison for raw locomotion. It is 180 cm tall, 47 kg, has about 2 hours of battery life, and is known for fast walking, running, stair climbing, and terrain adaptation. That is excellent for robotics labs and enterprise developers. It does not mean H1 is a plug-and-play home robot for a parent who wants laundry carried upstairs.

Unitree G1 is more accessible at a listed starting price of $13,500, but it is smaller at 132 cm and 35 kg, designed primarily for research and development, and does not carry the same home-stair claim in the database. Unitree R1 pushes affordability further, starting at $4,900 for the R1 Air pre-sale, with agile bipedal moves, push recovery, cartwheels, handstands, and a roughly 1-hour mixed-activity battery. That is remarkable pricing for a humanoid. It is also a reminder that acrobatics are not the same as certified household stair use.

1X NEO takes a different path. It is explicitly home-focused, soft-bodied, 167 cm tall, only 30 kg, and listed at $20,000 for early adopters, with roughly 4 hours of battery life. NEO's home-first design matters because human coexistence, softness, and chores are the real buyer problem. But stair autonomy is not the public proof point buyers should assume from that listing.

Other entries underline the trade-off. Boston Dynamics' Atlas Electric is an industrial humanoid with a 50 kg instant lift capacity, 90 kg body weight, IP67 rating, autonomous navigation, and dynamic recovery, but no normal consumer sale. RobotEra STAR1 claims outdoor terrain navigation, running, jumping, 55 degrees of freedom, and 20 kg payload, but again it is not a consumer staircase appliance. PrimeBot's Prime T1 is one of the few consumer-positioned robots explicitly framed around a transformable wheeled-humanoid and quadruped mode for stairs, slopes, and uneven terrain, but it remains a prototype with no announced price or detailed public specs.

Why stairs are harder at home than in a lab

A staircase is a hostile test environment hiding inside a familiar object. Humans ignore it because we have decades of embodied practice. Robots have to make the whole loop explicit.

The first problem is geometry. A robot has to understand riser height, tread depth, nosing, slope, landing shape, and whether the next surface is actually a step or a shadow. Even small errors matter because a bipedal robot is balancing a tall body over small contact patches.

The second problem is perception under bad conditions. Bright sun through a window, dark carpet, glossy wood, patterned runners, reflective edges, or a pile of shoes can all degrade depth estimates. Figure's stereo-camera framing is therefore meaningful, but buyers still need evidence across lighting and surface conditions, not one clean staircase.

The third problem is whole-body coordination. A humanoid cannot treat legs, torso, arms, and payload as separate systems. Carrying a laundry basket blocks cameras, shifts balance, and changes recovery behavior. Opening a door at the top of stairs adds manipulation and foot placement. Reaching for a handrail may help stability, but now the robot needs safe arm contact and force control.

The fourth problem is safety around humans. A 61 kg Figure 03, a 70 kg Unitree H2, or a 90 kg Atlas is not a toy. A staircase is narrow, elevated, and shared. A useful home robot must know when to wait, when to back down, when to ask for help, and when a child or pet makes the route off-limits.

The fifth problem is support. If a robot refuses a staircase, who diagnoses it? If an OTA update improves one stair pattern but worsens another, how is that tracked? Figure's production post mentions diagnostics, fallback ladders, fleet management, over-the-air updates, and field-service loops. Those are not boring enterprise details. They are the machinery needed before robots can be trusted outside scripted demos.

Are quadrupeds better for stairs than humanoids?

For pure mobility, often yes. Quadrupeds have four contact points, lower centers of mass, and a longer history of rough-terrain deployment. The ui44 database has several examples that make humanoid stair hype look less special.

Unitree As2, a mid-size quadruped, is listed with the ability to climb 25 cm stairs, traverse about 40-degree slopes, operate from -20°C to 50°C, and run over 4 hours unloaded. AGIBOT D1 Pro is much cheaper at $3,200 and is recorded with stair climbing up to 16 cm steps, 40-degree slopes, a 3.5 m/s top speed, self-balancing, anti-fall behavior, and 1-2 hours per charge. Roborock's Saros Rover is not a humanoid at all, but its wheel-leg architecture is explicitly designed to climb and clean stairs in multi-storey homes.

Those robots clarify the humanoid trade-off. If the job is only to traverse stairs, a quadruped or specialized wheel-leg robot may be more stable. The reason humanoids remain interesting is not that two legs are the easiest way to climb. It is that human homes are built around human reach, human tools, human stairs, human appliances, and human-scale furniture. A humanoid that can climb stairs and then use both hands upstairs is solving a broader home-access problem than a mobility-only platform.

That is also why the comparison page can be more useful than any single viral clip. The buyer question is not "which robot looked coolest on a staircase?" It is "which robot combines mobility, manipulation, battery life, safety, support, and availability in a way that matches my home?"

What proof should buyers ask for?

Do not ask only whether a humanoid can climb stairs. Ask what kind of evidence backs the claim.

The best public proof would include uncut runs across different staircases, lighting conditions, and floor materials. It would disclose attempt counts, not just successful clips. It would show the robot stopping when a person blocks the path, recovering from a misstep, refusing a staircase it cannot parse, and continuing safely after carrying a realistic object.

Battery and payload should be tested together. Empty-handed stair climbing is useful, but a home helper will be asked to carry laundry, small packages, cleaning supplies, or assistive items. Carrying changes center of mass, camera visibility, arm availability, and fall risk. If a robot has a 20 kg payload on paper but cannot use stairs while carrying anything safely, the payload number is less relevant for a multi-level home.

Support terms should be concrete. A company should explain whether stair use is approved, experimental, remotely supervised, restricted to mapped staircases, or forbidden by the warranty. It should provide emergency-stop behavior, logs, fleet update policy, service response, and a clear human override. The more powerful the robot, the less acceptable vague "AI will figure it out" language becomes.

Should stair climbing change your home robot buying plan?

Not yet, unless you are buying for research or development. For normal buyers, stair climbing should change how you evaluate future robots, not push you into a purchase today.

If you want a home humanoid now, 1X NEO is the clearest home-positioned pre-order in the database, but its value case is home coexistence and chore learning, not proven autonomous stair service. If you want a low-cost bipedal research platform, Unitree R1 and Unitree G1 are fascinating, but they are not appliance-like household helpers. If you want serious legged mobility, quadrupeds such as Unitree As2 or AGIBOT D1 Pro may be more immediately relevant, though they do not give you humanoid arms for upstairs chores.

Figure's System 0 claim is still worth watching closely. It combines the right ingredients: perception, whole-body control, simulation-trained reinforcement learning, fleet scale, diagnostics, OTA updates, and a manufacturing ramp large enough to generate real failure data. If those ingredients hold up in messy, public, repeatable environments, stairs could become one of the most useful benchmarks for home humanoid readiness.

For now, the safest conclusion is simple: humanoid stair climbing is becoming credible, but home stair use is not solved. The demo asks the right question. The product proof still has to arrive.