Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Learn how to train a quadruped robot to walk using reinforcement learning, from defining actions and observations to designing rewards and transitioning from simulation to reality.
| Date | Stars |
|---|---|
| 2026-07-25 | 290 |
| 2026-07-28 | 290 |
| 2026-07-30 | 290 |
| 2026-07-31 | 291 |
| 2026-08-02 | 292 |
| 2026-08-04 | 293 |
| 2026-08-06 | 293 |
Today
— stars today
This week
+3 stars this week
This month
— stars this month
Momentum
0.0
growth rate 1.03%/day
## Making Quadrupeds Learning to Walk: From Zero to Hero
<div><br>
<img src="./images/sim.gif" alt="Global Trajectory" style="width:80%; height:auto;">
</div><br>
Mobile robotics has reached a huge turning point. Thanks to the development and improvment of parallel computation and deep learning, robots are now able to learn complex tasks such as walking, running and jumping. The applications are endless, such as: ispection, search-and-rescue missions, entertainment and even healthcare.
>**Tasks that were once considered impossible for robots are now within reach, and the possibilities are endless. It's just a matter of time before we see robots performing tasks that were once only possible in science fiction.**
In this article, we will guide you through the process of training a quadruped robot to walk using reinforcement learning. We will cover the control policy, actions, observations, reward design, training process, and the transition from simulation to reality. By the end of this guide, you will have a solid understanding of the key components involved in training a quadruped robot to walk!
## 1. Problem Overview
<div><br>
<img src="./images/rl-framework.png" alt="Global Trajectory" style="width:60%; height:auto;">
</div><br>
In autonomous systems, a fundamental task is tracking reference commands. Usually a high-level controller provides linear and angular speed references, which the robot must follow while maintaining stability and preventing falls. Our final goal will be to track reference speeds generated by a user moving a joystick, while maintaining a desired height and avoid falling.
To achieve this, the robot will have to learn a control policy that maps observations to actions. The observations will include the robot's internal states and external inputs (i.e. user commands), while the actions will be the motor positions that the robot needs to achieve to move accordingly.
This is a really complex task, as the policy has to learn how to balance the robot, move the legs in the right way, and track the reference speeds. This is why reinforcement learning is a good approach for this task, as it allows the robot to learn from its own experience and improve over time.
## 2. Ingredients and RL Framework
A reliable control policy lies at the heart of any robotic locomotion system, acting as the decision-making mechanism that translates high-level objectives into low-level motor commands. The first step is selecting a suitable environment, where the primary focus is on replicating physical constraints while allowing for iterative training. In many cases, this environment can be a simulator equipped with detailed robot models, such as those derived from URDF files, which define the geometry and dynamics of the robotic system. By running experiments in a controlled and often accelerated environment, researchers can quickly gather feedback on different policy configurations while reducing the risk of hardware damage. The simulator can also be augmented with sensor data and randomization factors to expose the policy to varied conditions, promoting robustness that will extend beyond the simulator.
<div><br>
<img src="./images/policyaa.gif" alt="Global Trajectory" style="width:70%; height:auto;">
</div><br>
Once the environment is established, it is critical to consider how states, actions, and rewards are designed for effective learning. The state typically includes proprioceptive information, such as joint angles and velocities. Actions correspond to the motor commands needed to drive the robot’s joints, which the policy must produce in response to incoming observations. The reward function is designed to incentivize stable and efficient locomotion, often emphasizing forward progress, energy efficiency and reduced body tilt or impact forces. By balancing these elements, the policy gains a clear sense of what behaviors are desirable and how to achieve them through repeated interaction with the enviroExcerpt of 23,697 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:fa7329b8cc73c742, topic:robotics, topic:simulation, readme:robotics
matched fp:fa7329b8cc73c742, topic:reinforcement-learning, desc:reinforcement learning, readme:reinforcement learning