Robot dog learns to switch gaits like animals, tackling forests and stairs with new AI framework

Robot dog can climb stairs, navigate a forest and bound over logs thanks to new, rapid AI training technique

Researchers used reinforcement learning to train a quadrupedal robot to adapt to different environments using two different pre-learned gaits. A four-legged robot has learned to change the way it runs while navigating forests, staircases and obstacle courses. -- seamlessly switching between a steady trot and a faster bounding gait without instructions from a human operator. The 100-pound (45 kilograms) robot, called KAIST HOUND, uses cameras and lidar to scan the ground ahead, then selects an appropriate gait and adjusts its movements in real time. In outdoor tests, it crossed a 0.7-mile (1.1- kilometers) university campus route and a 0.2-mile (0.3 km) forest trail strewn with roots, logs and slippery leaves. The researchers described the robotic framework on July 15 in the journal Science Robotics. Changing gait Animals naturally change their gait depending on their speed and surroundings. A dog might trot carefully across uneven ground, for example, before bounding over a fallen branch. Reproducing this adaptability in robots is tricky because different movements are often controlled by separate, highly specialized coding systems, and transitions between them can cause a lag that drives the robot to stumble. To overcome this issue, researchers developed a special training framework called action pretrained transformer-based reinforcement learning (APT-RL). This is an artificial intelligence (AI) training system that first studies many examples of actions, uses a transformer to understand patterns across those actions, and then improves through rewards and penalties. The training began with a simple, two-dimensional computer model of the robot. Using trajectory optimization -- a technique that calculates physically workable movements for the robot -- the team generated 180,000 short trotting and bounding sequences, including the joint forces the robot's legs need to perform. The dataset represented about 15.5 hours of movement but took only around eight minutes to produce. During reinforcement learning -- a machine learning technique where AI learns to make the best decisions by engaging with a particular environment through trial and error -- an AI system then learned how to select and modify those skills while negotiating simulated stairs, stepping stones, hurdles, gaps and rough ground. In digital simulations, the robot dog was not limited to copying its prerecorded movements. It could also make corrections for three-dimensional terrain and unexpected situations, such as jumping over a log -- a behavior that wasn't included in the original, flat-ground training data. Finally, the researchers configured the system to include the robot's depth camera and lidar scanner in the simulation. In one indoor test, HOUND bounded across an obstacle 2 feet (60 centimeters) high while briefly achieving 9.5 mph (15 km/h). It also jumped down a three-step staircase. The robot generally chose trotting at lower speeds on irregular ground, while bounding became more common at higher speeds or when it encountered larger steps, hurdles or gaps. The AI system that could select either gait performed more consistently across the different simulated environments than the version restricted to trotting or bounding alone. The researchers suggest the technology could eventually help robots navigate disaster zones or other places inaccessible for wheeled machines. However, the current framework only allows two gait choices and mainly handles forward movement. Rapid turning, sideways motion and other behaviors like crawling remain future goals for the research team.

Robot dog learns animal-like movement to tackle forests, rocky trails

Researchers from Korea have developed an AI framework that enables a four-legged robot to autonomously switch between different motor skills while moving across complex terrain. The system allows the robot to adapt its gait in real time to navigate forests, climb stairs, and jump over obstacles using only onboard sensors and computing. The approach combines pretrained locomotion skills with adaptive decision-making to improve agility in changing environments. The researchers say the framework could support future autonomous search-and-rescue and exploration missions. "A single onboard policy enables robust traversal of diverse obstacles, including stairs, hurdles, stepping stones, gaps, and fallen branches, demonstrating the versatility and effectiveness of our approach," said the team in the research paper. Researchers from the Korea Advanced Institute of Science and Technology (KAIST), Korea University, the Agency for Defense Development, and DIDEN Robotics have developed Action Pretrained Transformer-based Reinforcement Learning (APT-RL), a new AI framework that enables four-legged robots to automatically switch between different movement skills while navigating complex terrain. Unlike conventional legged robots that rely on a single gait or separate controllers for different tasks, APT-RL allows the robot to seamlessly transition between trotting, bounding, climbing, and jumping based on the terrain ahead. The system combines pretrained motor skills with reinforcement learning, enabling the robot to choose the most suitable movement strategy in real time using only its onboard sensors. The framework trains the robot in three stages. First, it learns basic locomotion skills from more than 180,000 optimized motion trajectories, representing 15.5 hours of simulated movement generated in just eight minutes through trajectory optimization. These motion patterns form reusable building blocks for more advanced behaviors. Next, reinforcement learning teaches the robot when and how to switch between these skills while maintaining balance across different obstacles. Finally, the system is adapted for real-world deployment using only onboard perception from a depth camera and a 2D LiDAR sensor, eliminating the need for external motion-capture systems. The researchers say the framework enables faster, more adaptable locomotion across unpredictable environments and could support future search-and-rescue, exploration, and other autonomous robotic missions. The framework was tested on KAIST's quadruped robot, HOUND, in both indoor and outdoor environments. The robot successfully navigated university campuses, forest trails, grassy fields, rocky terrain, staircases, stepping stones, logs, gaps, and fallen branches. It automatically selected different gaits depending on the terrain and its commanded speed. For example, it used a stable trot while negotiating uneven ground or climbing stairs, then switched to a faster bounding gait when jumping over logs or descending large steps. The robot demonstrated impressive speed as well. During obstacle traversal, it reached an instantaneous speed of 4.25 meters per second. While jumping down a three-step staircase, it briefly reached 6 meters per second before landing. According to the researchers, these are among the fastest reported speeds for a perception-driven quadruped robot operating in complex real-world environments. The researchers also compared their approach with existing reinforcement learning and hierarchical control methods. APT-RL consistently achieved better success rates, smoother gait transitions, improved energy efficiency, and faster learning. The robot adapted to previously unseen terrain without requiring retraining, showing that the pretrained movement skills could be reused in new situations. The study also found that combining a depth camera and LiDAR produced the best performance. The depth camera provided detailed information about nearby obstacles, while the LiDAR allowed the robot to detect terrain several meters ahead, giving it enough time to plan its movements at high speed. The researchers believe the framework could eventually support autonomous robots operating in forests, disaster zones, construction sites, or other challenging environments where wheeled machines struggle. Future work will expand the system to include additional gaits such as galloping and crawling, support more agile turning and sideways movement, and adapt the same learning framework to humanoid robots and other legged machines.