NVIDIA Uses AI Agents to Help Microduck Learn to Walk in Simulation

On October 23, 2026, NVIDIA introduced simulation using AI agents as a way to teach robots such as Microduck to walk. The article explains that a walking policy alone is not enough; robots also need to perceive their surroundings, maintain balance and adjust their actions.

NVIDIA's explanation shows that learning to walk is not a task for a single control model but a complex one involving perception and balance adjustments. For developers, repeatedly training robots on obstacle courses in simulation, rather than testing them by risking damage to real machines, may be worth considering.

The new aspect of this approach is its emphasis on training robots not just to walk as a standalone control problem, but to use AI agents to learn environmental understanding and action adjustment together. According to NVIDIA, robots need to perceive their surroundings, maintain balance and correct their movements, in addition to following a walking policy. The training environment is therefore not simply upright walking but a complex setting such as an obstacle course. In other words, the focus here appears to be less on the robot's physical movement itself than on a learning framework that allows it to adapt its behavior to circumstances in simulation. No related coverage from this newspaper was cited, so it is not possible to connect this to earlier reporting. However, based on the information made public, elements shown in the image, such as ramps, platforms and rubber mats, represent conditions with much greater uncertainty than a level indoor environment. This suggests the approach may be relevant when testing not only walking accuracy but also adaptation to steps and changes in flooring. The question here is not so much whether simulation can replace real-world testing as how much more intensive training can be conducted in simulation before deployment on real machines. The broader industry implication is that competition in robotics development appears to be shifting from hardware alone toward learning platforms that incorporate the environment. The statement that a walking policy is insufficient indicates that mobile robots need multiple types of input, including images, posture and ground contact, making simulation design itself a potential source of performance differences. Unresolved questions include how broadly this approach can generalize across robot types and terrain, where errors remain when transferring to real machines, and what kinds of behavioral policies AI agents generate during training.

The published article cites Microduck as an example of a robot navigating an obstacle course and shows an environment with ramps, platforms and rubber mats. NVIDIA says training robots in such environments requires complex learning that includes perception and posture control, rather than walking control alone.

For Japan's robotics developers, simulation design that accounts for flooring and steps, in addition to walking control, may become more important. A key question, particularly for mobile robots and humanoids, is how thoroughly training environments can be built before moving to real-world testing.