Humanoid Robot G1 Learns to Climb Stairs and Traverse Rough Terrain: A Single Policy Controls Nine Body Parts

Humanoid robots traversing complex terrain have long been limited by a single locomotion mode: "walking only." A recent study proposes a very different approach: let the robot actively use multiple body parts for support and propulsion, enabling it to climb stairs, crawl, and vault over obstacles. This work, built on the Unitree G1 platform, achieved a 92.5% success rate in simulation, but leaves a key question mark—whether this flexibility will hold up equally well in the real world.

Core Breakthrough: From "Bipedal Walking" to "Whole-Body Support"

Traditional humanoid locomotion control is typically built around bipedal gait, with the arms and torso used mainly for balance or manipulation. What sets this study apart is that it trains a single control policy that lets the robot autonomously decide where and how to apply force across nine different body parts. These nine parts include both hands, both elbows, both knees, both feet, and the pelvis.

In other words, the robot no longer treats "walking" as the only solution. Faced with stairs, it can climb using both hands and feet; faced with a low space, it can crawl through; faced with a high ledge, it can brace against the edge with its hands and vault over; and even after a fall, it can push itself back up using its elbows and knees. This strategy of "use whatever can provide leverage" is closer to human instinct in complex environments.

Why a Single Policy Matters

A common approach for multi-skill robots is to train a separate policy for each task and then dispatch among them via a switching mechanism. This approach carries high engineering complexity, and transitions between tasks are prone to failure. If a single policy can, under one set of parameters, cover climbing, crawling, vaulting, stepping-stone walking, and fall recovery, it means the controller's generalization boundary has been significantly widened. For real-world deployment, this reduces the burden of separately tuning and maintaining parameters for each terrain type.

Single policy versus multi-policy switching are two typical architectures in robot learning. Multi-policy solutions usually train separate neural network controllers for tasks such as climbing, walking, and recovery, then rely on a high-level scheduler to switch based on terrain recognition results. This hierarchical design is simple to implement but has two inherent flaws: first, abrupt motion changes or control vacuums tend to occur at switching boundaries; second, the observation and action spaces of different policies are often inconsistent, making it hard to share experience. A single policy requires one network to cover all behaviors under the same set of input-output definitions, and its training typically relies on large-scale parallel simulation and curriculum learning in reinforcement learning, gradually increasing terrain difficulty. The cost is a significant rise in training difficulty, because the optimal exploration paths for different skills may conflict with one another; the benefit is that implicit transitions between skills form within the policy, reducing the fragility introduced by explicit switching.

Performance and Concerns in Simulation

The research team validated the policy on a simulated obstacle course, achieving an overall success rate of 92.5%. This figure is quite impressive in simulation, but the original text also clearly states that the bigger test is whether this flexibility can transfer to unpredictable real-world environments.

Simulation can precisely control friction, lighting, and terrain parameters, whereas the real world is full of material variations, slippage, sensor noise, and dynamic disturbances. A policy that can reliably brace against a step with its hands in simulation may fail on real stairs because a handrail is loose or the surface is wet and slippery. The 92.5% success rate is proof of capability, but not a guarantee of deployability.

New Risks Introduced by Whole-Body Support

Using parts such as the pelvis and elbows to support the body also introduces new contact risks and wear issues. These parts may not have been designed for high-frequency load bearing, and their long-term reliability and safety require additional evaluation. Moreover, multi-point contact demands higher force-control precision, and a failure at any single contact point can trigger a cascading loss of balance.

Force-control precision in such whole-body contact tasks refers to the robot's ability to regulate the normal force and tangential friction force applied at each contact point. Traditional bipedal walking focuses mainly on foot contact, with relatively few force-control dimensions; when the pelvis, elbows, and knees all participate in support simultaneously, the controller must coordinate force distribution across multiple contact points to avoid overloading or slipping at any one part. This typically relies on jointly solving with torque sensors and a whole-body dynamics model, and is extremely sensitive to perception latency and modeling errors. At the same time, structural durability cannot be ignored: the elbow and knee joints of humanoid robots are mostly designed for swinging motion, and the rated loads of their reducers and linkages are not intended for continuous load bearing. High-frequency bracing on the ground with elbows and knees can lead to joint overheating, lubrication failure, and even structural fatigue—factors that are usually simplified as rigid contact in simulation and thus cannot be represented.

Sim-to-real transfer is the core bottleneck for deploying robot learning. To transfer a policy trained in simulation to a physical robot, domain randomization is usually required—randomly varying parameters such as friction coefficients, mass distribution, sensor noise, and actuator delay during training so that the policy is robust to real-world uncertainty. But domain randomization can only cover the types of disturbance the designer anticipates; long-tail situations in the real world, such as loose handrails, slippery floors, and sudden lighting changes, may still fall outside the training distribution. In addition, whole-body contact tasks demand far higher contact-force feedback precision than pure legged walking, while the torque-sensing noise and joint compliance of real robots often have systematic deviations from simulation models. Such deviations are amplified during multi-point contact and become a major source of transfer failure.

What This Means for the Industry

The value of this work lies not in the single capability of letting a robot "climb a ladder," but in demonstrating a control paradigm: treating the whole body as callable contact resources rather than completing tasks solely with end effectors. For scenarios such as inspection, rescue, and home service that need to cope with unstructured environments, this approach offers a new technical path.

Using the Unitree G1, a relatively lightweight humanoid platform, to validate such whole-body control policies also shows that the practicality of small and mid-sized humanoid robots in algorithm research is increasing. Falling hardware costs and a maturing open-source ecosystem are enabling more teams to reproduce and advance such work.

However, from simulation to reality and from demo to product, there are still many engineering problems to solve. The robustness of whole-body contact policies, the structural durability of contact parts, and the perception closed loop in real environments are all key variables determining how far this path can go.

This article is based on Reddit community discussion and publicly available research information, with core data cited from the original report.

One-line Summary

A single policy lets a humanoid robot use nine body parts for leverage and support, achieving a 92.5% success rate in simulation, with real-world transfer still to be validated.

Paragraph Summary

Using the Unitree G1 platform, this study proposes a control paradigm different from traditional bipedal gait: training a single control policy that lets the robot autonomously decide where and how to apply force across nine body parts—both hands, both elbows, both knees, both feet, and the pelvis—to climb stairs, crawl, vault over obstacles, walk on stepping stones, and push itself up after a fall. Compared with training separate policies for each task and switching among them, a single policy significantly widens the controller's generalization boundary and reduces deployment and maintenance costs. On a simulated obstacle course, the policy achieved a 92.5% success rate. However, the article notes that the precise parameter control of simulation cannot represent real-world material variations, slippage, and sensor noise, and whole-body contact also brings new risks in force-control precision and the structural durability of parts such as elbows and knees. The value of this work lies in demonstrating a control paradigm that treats the whole body as callable contact resources, offering a reference for unstructured scenarios such as inspection and rescue.