Tencent's MAX Quadruped: Wheels Meet Legs in Perfect Harmony

Introduction to MAX - the Multimodal Quadruped Mobile Robot

Is it a legged robot, or is it a wheeled robot? Well, it’s both. But why, you ask? Why not! On flat surfaces, MAX can speed along using its wheels, and when faced with challenging terrain, it switches to its legs to move with excellent agility. Ingeniously, the wheels are incorporated into the robot’s knees. By bending its knees, the wheels are activated, transforming MAX into a wheeled mobile robot ready for fast locomotion.

If you've seen the videos, you’ve probably noticed that MAX can perform an impressive range of movements: standing upright on two wheels, jumping over obstacles and dense plum blossom piles, balancing on a single pile with a small foot area, performing squats and somersaults, and freerunning over various obstacles.

These movements aren’t easy. The robot encounters shocks, bumps, shaky terrain, and more, requiring constant adaptation and adjustment. Nonetheless, Max executes the moves with precision and agility. And impressively, MAX's positioning accuracy is less than 1%, and its terrain recognition accuracy is within 2 cm.

Let's dive deeper into what makes MAX so incredible!

What Makes MAX Quadruped Special?

To perform its impressive array of movements, MAX is equipped with advanced intelligence and a suite of sophisticated algorithms. For instance, it uses perception algorithms to observe and understand its environment, recognizing terrain and obstacles. This allows MAX to perform intricate trajectory and motion planning (such as omnidirectional six-degree-of-freedom motion planning) to navigate its path and movements in real-time.

MAX also relies on control algorithms, like model predictive control, to manage its movements and prevent failures such as falling. These algorithms enable MAX to adjust its actions dynamically, ensuring stability and precision.

But these are just a few of the many algorithms integrated into MAX. It’s continually undergoing enhancements and optimizations, both in terms of algorithms and design, making it more capable and efficient with each iteration.

Incorporating Artificial Intelligence for Improved Agility

MAX Quadruped also boasts advanced artificial intelligence capabilities, including self-developed model algorithms and deep reinforcement learning. By integrating cutting-edge pre-trained AI models and reinforcement learning technology into its control systems, MAX's flexibility and autonomous decision-making abilities have been significantly enhanced. This enables the robot to adapt and respond to complex and changing environments with ease, further showcasing its remarkable agility and intelligence.

Learning in Stages

Tencent Robotics X Lab has developed a three-stage learning process for their robot dogs using pre-training models and reinforcement learning. Summarized briefly, in the first stage, they collected real dog motion data, like walking and jumping, and used this to train the robots to imitate these movements. In the second stage, the robots were taught to link these movements with environmental data, allowing them to adapt to different surroundings. Finally, in the third stage, the robots used their accumulated knowledge to solve complex tasks end-to-end by analyzing information and making top-level strategic decisions. This progressive learning allows the robots to build on their knowledge continuously and respond flexibly to new challenges.

Training Using "World Chase Tag"

To train MAX's new skills, researchers created an obstacle chase game inspired by "World Chase Tag," where two athletes compete, one as the pursuer and the other as the evader. In the robot version, two MAX robots are placed in a 4.5 x 4.5-meter field with obstacles. One robot is randomly assigned to pursue while the other evades and aims to touch a flag. If the evader touches the flag, roles switch, and the game continues until the pursuer catches the evader.

This game helped MAX develop reasoning and decision-making abilities using deep reinforcement learning. For instance, if the pursuer realizes it can't catch the evader before it reaches the flag, it will wait for the next opportunity. The robots also exhibit behaviors like jumping to catch the evader or evading obstacles, mimicking real-world animal actions. These strategies, learned in simulation, transfer seamlessly to real-world scenarios, enhancing MAX's ability to adapt and perform complex tasks autonomously. But this is just one way to train the robot dogs, similar games can be utilized to enhance the learning process.