Watch the dog sail from sidewalk to forest floor – notice it trots to the boundary, adjusting its gait without consciousness. Now, scientists at the University of Leeds have taught a four-legged robot, Clarence, to do the same thing, learning teenage animals for days or weeks in just nine hours.
The robot hasn’t done anything machine-friendly: autonomously switch between eight different gaits (stepping, running, boundary, jumping, etc.), purely under the terrain conditions encountered. Unlike robots that currently have to be programmed for specific motions, Clarence can adapt to its big strides in real time, even on surfaces that have never been experienced.
Animal blueprint
Animals switch gaits to survive to save energy, maintain balance or escape predators. A horse goes from walking to galloping by following the rulebook, but by strategies embedded in its nervous system. The Leeds team reversed these biological principles into artificial intelligence.
“Our findings could have a significant impact on the future of leg robot motion control by reducing many previous restrictions on adaptability,” noted graduate researcher Joseph Humphreys. His framework embeds three core animal abilities: gait transition strategy, program memory of different movements, and real-time motion adjustment.
The breakthrough is to teach the robot not only to move, but also how to decide which movement to use. Traditional robots follow pre-programmed patterns – symbolizing attempts to navigate various terrains while locking in a single walking style. Clarence learned to evaluate conditions and choose the best gait at any time.
Beyond Programming: Learning Choice
The research team developed what they call biologically inspired indicators—mathematical representations of the principles used for gait selection in animals:
- Energy efficiency: Minimize energy and maintain energy
- Stablize: Maintain balance between unpredictable surfaces
- Force Management: Protect joints and actuators from overstrain
- Work optimization: Reduce the mechanical effort required to exercise
Together these indicators work – as in nature, a single factor can drive gait selection. When Clarence encounters loose wood, the sensor detects instability, and the AI quickly turns from trotting toward the boundary, restoring balance before a catastrophic failure.
Real-world verification
The real test is outside the lab. The researchers released Clarence on terrain that they had never seen before: muddy grass, piles of rocks, overgrown roots, and even moving wood under their feet. The robot navigates them all and automatically switches gait according to the conditions.
A surprising parallel is the behavior of wild animals. When Clarence encountered particularly challenging terrain, it deployed assisted gait, the specific movement and lim used by the animals for stability recovery. Researchers have not yet programmed this strategy; it comes naturally from the decision-making process of AI.
Professor Zhou of UCL Computer Science explains the importance: “Instead of training robots to target specific tasks, we want to use strategic intelligence for animals to adapt to gaits to use principles such as balance, coordination and energy efficiency.”
From simulation to reality
Training uses deep reinforcement learning entirely in virtual environments – basic high power trials and errors on hundreds of simulated terrain. This approach reflects how Neo learns martial arts in the matrix, as Humphreys points out: “All training is done in simulations. You train policies on the computer, then take them and put them on the robot, and you are as proficient as the training.”
What makes this striking is a seamless transfer from simulation to reality. Despite never experiencing rough terrain during training, Clarence succeeded in leading to a complex real-world surface on his first attempt, a notorious challenge in robotics known as the “SIM to SIM to Real Gap”.
The robot achieves 90.6% accuracy in gait selection while selecting the least amount of energy, which is only 4 joules, completing tasks that usually require more power. This efficiency is critical for applications where battery life determines the success of tasks.
What goes beyond robotics
This framework opens avenues for robots in dangerous environments where humans have risks: nuclear decommissioning, disaster response, planetary exploration. Current robots often fail when encountering unexpected conditions – this limitation can be fatal in rescue situations.
Perhaps more interesting is that this study provides a new tool for studying animal biomechanics itself. Instead of burdening the live animals with invasive sensors or dangerous experiments, researchers can use robotic alternatives that replicate natural motion patterns to test hypotheses.
This work represents a fundamental shift from programming specific behaviors to inculcating adaptive intelligence. As artificial systems become more autonomous, the biological principles that allow millions of years of successful navigation may prove to be a valuable guide to the next generation of adaptive machines.
Related
Discover more from Neuroweed
Subscribe to send the latest posts to your email.
