Sony’s Aibo just got a stealth upgrade thanks to AI.
Researchers from ETH Zurich and Sony created an AI-powered system to help the robot dog walk more quietly. The launched the study in response to common complaints that quadrupedal robots are too noisy for home use. In paper, they noted that technological advancements are driving interest in consumer robotics. The more advanced home robots get, the deeper the emotional connections get.
“For example, Aibo, a robotic pet introduced by Sony, is capable of learning and developing its own unique personality through interactions with its home environment and owner,” they wrote. “This emotional connection with robotic pets demonstrates the potential for robots to play a significant role in our daily lives.”
They developed a reinforcement learning-based AI model to teach AIBO to walk more softly in a virtual environment. Within Nvidia’s Isaac Simulator, the robot’s digital twin taught itself to reduce footstep noise through trial and error.
According to the researchers, three key techniques helped AIBO walk quieter:
- Changing joint stiffness to absorb impact.
- Using foot sensors to adjust how feet hit the ground.
- Training the robot in steps to slowly improve.
Testing proved that the reinforcement learning (RL) method achieved quieter results than Sony’s own walking controllers. They found that faster walking increased noise, but the AI model kept it lower than other methods at all speeds.
The new AI method reduced walking noise by about 5 dB compared to Sony’s built-in quiet mode and by 10 dB compared to standard AI-driven locomotion. The quietest result achieved was 30 dB. At that level, a human might hear the faint noise in a silent room if they listen closely.
They used a previous AI model trained for walking but not for quietness as a baseline RL policy. The robot’s built-in microphone recorded footstep noise while walking. The researchers focused on human-audible frequencies between 20 Hz and 20 kHz. Sound levels were recorded for 10 seconds for comparison.
However, there was a tradeoff. The quieter Aibo got, the more it struggled traversing rough surfaces. The researchers balanced quietness and stability by tweaking friction, terrain height, and other conditions in simulations.
The researchers said this is the first time AI has been used to reduce robotic footsteps in a quadruped robotic pet. The insights could prove useful as a growing number of companies race to mass market their quadruped and humanoid robots for households.