Researchers at ETH Zurich and Sony Group Corporation have developed RL models to help the robot walk quietly and dance expressively.
Sony’s Aibo, a robot puppy designed as a household companion, can mimic real dog behaviors like walking on four legs, responding to its name, reacting to toys, performing tricks, and responding to cuddles.
Researchers at ETH Zurich and Sony Group Corporation have developed new reinforcement learning (RL) models to enhance Aibo’s abilities. These models aim to help the robot walk quietly and perform expressive dance routines.
Users of the robot have noted the noise of its footsteps while moving around the home. To address this, researchers recently developed a simulation-based reinforcement learning approach to reduce foot contact velocity in a physics simulator, enabling quieter walking.
The team have developed an RL-based approach to reduce the noise Aibo makes while walking. Their method adjusts joint stiffness and damping in real time, using sensor data from the robot’s paws. It also applies penalties to rapid movements that generate noise, gradually encouraging quieter steps.
The researchers tested their approach on Aibo in real-world experiments and compared its noise levels to those produced using baseline RL methods and Sony’s existing controllers. The locomotion controller significantly reduces the noise produced by the robot while walking, making it much quieter than commercial and conventional RL controllers.
Beyond enabling Aibo to walk more quietly, the researchers aimed to enhance its entertainment abilities. They developed a model called Deep Fourier Mimic (DFM), which combines motion representation with RL to generate more expressive dance routines.
The team have demonstrated that their new dancing policy enabled smoother, more natural movements. It also allowed Aibo to interact with users in new ways, such as dancing alongside them and mimicking their motions.
These RF-based approaches could make Aibo quieter and enhance its entertainment capabilities. They might also be applied to other household robots or robotic systems designed to entertain guests at theme parks worldwide.
The team aims to address the limitations observed during real-world testing of their models in future work.
Reference: Ryo Watanabe et al, Learning Quiet Walking for a Small Home Robot, arXiv (2025). DOI: 10.48550/arxiv.2502.10983
Ryo Watanabe et al, DFM: Deep Fourier Mimic for Expressive Dance Motion Learning, arXiv (2025). DOI: 10.48550/arxiv.2502.10980
