Sony’s aibo dog could soon walk quietly and perform elaborate dance routines

Aibo, the cute-looking robot puppy developed by Sony, was designed to be a household companion. The robot can already emulate many of the movements and behaviors of real dogs, such as walking on four legs, responding when it is called by its name, reacting to toys, performing tricks and responding to cuddles.

Researchers at ETH Zurich and Sony Group Corporation recently developed new reinforcement learning (RL)-based models that could further expand the aibo robot’s capabilities. These models, outlined in two different papers pre-published on arXiv, are specifically meant to allow the robot to walk quietly and perform expressive dance routines, respectively.

“Users of aibo, the robot developed by Sony, have highlighted the noise of the robot’s footsteps as it is walking around their home,” Ryo Watanabe, first author of the studies, told Tech Xplore.

“We thus recently designed a sim-to-real based RL approach to minimize the foot contact velocity in the physics simulator to achieve quiet walking.”

To reduce the noise that the aibo puppy makes while walking, Watanabe and his colleagues developed a new approach based on RL. Their method was trained to actively dampen and stiffen each of the robot’s joints, leveraging data collected by the sensors under its paws and gradually enforcing penalties on rapid movements that produce noise.

The researchers applied their approach to aibo and assessed its performance in real-world experiments. They then also compared the noise that the robot made when employing their method to the noise it made when its locomotion was guided by other baseline RL-based methods and the controllers currently used by Sony.

“We found that our locomotion controller significantly decreases the noise produced by the robot while walking, making it significantly quieter than commercial controllers from Sony and other conventional RL controllers,” said Watanabe.

In addition to equipping aibo with the ability to walk more quietly around home environments, the researchers set out to enhance its entertainment skills. Specifically, they created a model called Deep Fourier Mimic (DFM), which combines motion representation with RL to produce more expressive dance routines.

“Creating diverse artistic motions for entertainment robots is time-consuming for designers and typically restricted to simple motion playback,” explained Watanabe. “In contrast, DFM allows entertainment robots to imitate an artistic motion as much as possible while adding an additional task, such as locomotion.”

Watanabe and his colleagues showed that the new dancing policy they created resulted in smoother and more natural movements. Moreover, it allowed the aibo robot to interact with users in new ways, dancing with them and mimicking their movements.

“The dancing policy we developed not only tracks reference motions but also transitions between different motion modes naturally,” said Watanabe.

The new RF-based approaches designed by this team of researchers might be applied to aibo, making it quieter and further improving its ability to entertain users. In addition, they could be applied to other household robots or to robotic systems designed to entertain guests at theme parks worldwide.

In the future, Watanabe and his colleagues hope to overcome some of the limitations they observed when testing the models in real-world experiments.

“We observed a trade-off between quietness and robustness during walking,” added Watanabe.

“We thus now aim to develop a method for learning a walking policy that is both quiet and robust by leveraging perception information. The expressive dance method, on the other hand, exhibited an advantage for periodic motions such as dance and locomotion, but not for non-periodic motions like grasping, kicking, and standing up, which currently need to be segmented separately.”

© 2025 Science X Network