09,05,2026 7 views
When a humanoid robot faces complex terrains such as plum blossom piles, broken bridges, and narrow stepped stairs, the real test that lies in front of it is far more than just "seeing under its feet".
On continuous flat ground, there is usually room for adjustment for an inaccurate landing; But in sparse, discrete, and geometrically limited landing areas, a deviation of a few centimeters from the sole of the foot may result in stepping empty. Once a robot makes a wrong choice in an instant, its body often loses balance before it can adjust.
A key question arises: can robots predict in advance what future this step will take them to before taking the current step?
To address this issue, the Sweet Potato Robot Algorithm team proposed the World Model Augmented VisualLocomotion, abbreviated as WM-LOCO.
WM-LOCO introduces a world model for future vision and motion state prediction into the complex terrain walking of humanoid robots, conducts end-to-end joint training in simulation, and achieves zero shot real machine transfer. This means that the strategies trained in simulation can be directly deployed to physical robots without the need for secondary adjustments.
WM-LOCO architecture diagram: PPO strategy on the left and cyclic world model on the right, sharing encoder and end-to-end joint training
Simply put, in the past, robots relied more on what they saw in front of them to determine their next steps; WM-LOCO has taught robots to "predict the future first and then decide what to do next".
1、 Why is it not enough to only understand the past?
Traditional visual reinforcement learning typically inputs current and historical visual information along with the robot's body state into a strategy network, and then outputs real-time joint control instructions.
This type of strategy based on observation history can "see" the information that the robot has already experienced. By inputting multiple consecutive frames of images or a segment of ontology state history, it can determine how the robot has previously moved and how the terrain has changed, and generate the current action based on this.
But the problem is that the strategy always utilizes' past style 'information and does not explicitly learn' the future state changes that a certain action may bring '. This limitation is not yet evident in flat and regular terrain, but once faced with terrain such as plum blossom piles, broken bridges, narrow tread stairs, etc. where landing areas are strictly limited, the shortcomings of decision-making based on observation history will be sharply magnified.
Because walking on complex terrain is never a simple superposition of unrelated single step actions.
Where the current foot is stepped on will continue to affect the body's center of gravity, support relationship, swing leg trajectory, and which areas can be reached in the next step. A seemingly feasible action at this moment may leave the robot in a state of no way out in the next moment.
So, robots not only need to understand 'what happened before', but also need to anticipate: what might happen next if this action is executed?
2、 Equip strategy with an understanding of the 'future'
WM-LOCO simultaneously receives robot onboard deep vision and proprioceptive information, and continuously learns the dynamic relationship between robots, actions, and environment in hidden space through a cyclic world model.
It should be emphasized that the purpose of this world model is not to generate a future video for human viewing, but to learn more compact and suitable representations of future states for real-time control. WM-LOCO attempts to answer three questions at every moment:
What is the current state of robots and terrain?
How will the visual and physical states change after performing an action?
Is this change beneficial for stable walking in the future?
In practical implementation, the world model consists of a deterministic memory trajectory and a random latent variable: the former accumulates history along a time loop, while the latter infers from current observations.
By learning observation reconstruction, potential state transitions, and task feedback prediction, the world model gradually extracts dynamic representations containing future motion trends and feeds this information back to the PPO strategy.
Therefore, the strategy no longer only obtains steps, obstacles, and landing areas in the current depth image, nor is it just a historical motion trajectory, but also contains forward-looking information on how the robot's body state may evolve in the future.
GT represents the actual depth of the next frame, Recon represents the model prediction, and Error represents the pixel by pixel error
This is the core difference between WM-LOCO and ordinary multi frame input strategies. Observing history mainly helps strategies understand 'what happened in the past'; The world model further learns the dynamic relationship between observations, actions, and future states, enabling strategies to infer what may happen next.
As a result, strategic decision-making no longer relies solely on historical information, but truly possesses predictive capabilities for the future.
3、 End to end training, bidding farewell to tedious processes
In order to enable robots to pass through complex terrain, many technological routes will split the system into multiple modules such as perception, foothold planning, trajectory generation, teacher strategy, and low-level control.
Some plans also require training a "hanging" teacher strategy with complete terrain truth values first, and then transferring the ability to student strategies using airborne sensors through knowledge distillation. The cost of such schemes lies in longer training links, more manual design, higher data construction costs, and the possibility of errors between multiple modules being transmitted and amplified layer by layer.
In contrast, WM-LOCO has chosen a more direct and concise path: the world model and walking strategy are jointly optimized end-to-end in the same training process.
The entire system does not require manual annotation of landing points, does not rely on pre planned foot trajectories, does not require teacher student multi-stage distillation, does not require complex multi expert structures, and does not require training multiple independent modules separately. From visual input to joint movements, training is done in one step.
The dynamic information learned by the world model is not reconstructed for the sake of reconstruction, but directly serves a goal: to help strategies make better next decisions.
4、 Higher success rate, more natural gait
Under the same visual input and task settings, the algorithm team compared WM-LOCO with the baseline PPO strategy.
As the landing area becomes increasingly sparse, narrow, and discontinuous, there is a clear gap between the two methods. The pure PPO strategy that relies solely on observation history is more prone to problems such as premature landing, frequent probing, inconsistent left and right leg rhythms, and feasible local movements that cannot be continued in the future. In some high difficulty terrains, it is even difficult for PPO alone to form stable passability.
After introducing the world model, the success rate of robots in complex scenarios such as plum blossom piles, broken bridges, narrow tread stairs, and single plank bridges has significantly improved, approaching 100% in simulation. At the same time, the robot reduced abrupt corrections and hesitant gait.
In other words, the world model not only brings about the ability to walk over, but also enables robots to exhibit more continuous and stable movement intentions while walking - swinging legs, landing, and transferring body center of gravity are more coherent, left and right leg rhythms are more coordinated, and overall movement is more natural and closer to human walking.
On the real machine, the same set of weights was deployed to Unitree G1 in a zero shot manner: 10 tests were conducted on three types of terrain: stepping stones, steps, and gaps, with an average success rate of 93.3%. Among them, the gap of 0.8 meters wide was crossed by the robot in a complete "swing leg crossing landing" motion step by step. For safety reasons, the team did not continue to test wider gaps, but this is already a rare real machine crossing distance in current public work.
5、 Real machine on-site verification: not just demo
The commonly displayed demo videos on the market often fail to fully reflect the stability of algorithms in real environments after screening.
The on-site display is completely different. Robots need to face real sensor noise, device status changes, lighting interference, and single operation that cannot be restarted. Any perception or control error will be directly exposed to the audience.
At the recent Xuri S600 media communication conference, a humanoid robot equipped with WM-LOCO completed a public real machine display on site. The algorithm runs in real-time on real robots, real sensors, and the Xuri S600 computing platform, completing a complete closed-loop from visual perception, world modeling to full body motion control.
Compared with the complex terrain walking demonstration mainly presented in the form of video demos, the public on-site operation puts higher requirements on the stability, real-time performance, and deployment reliability of the algorithm, further verifying the engineering reliability, real deployability, and high success rate of WM-LOCO.
The world model is not only used to generate the future, but can also help robots understand the future.
When robots move from making decisions based on observation history to actively predicting the consequences of actions, their walking logic also changes: first predict the future, and then take the next step.
(excerpt)
