The aibo robopod will learn to dance and walk almost silently

However, a team of researchers from ETH Zurich and Sony Corporation recently presented two reinforcement learning (RL) approaches that could expand Robops' arsenal. The first method is designed to reduce walking noise, while the second is designed to master smooth "dance" movements.

What is the problem with noise?

Many aibo users complain that the robot bangs its paws loudly against the floor when moving around the house. A team led by Ryo Watanabe solved this problem using simulation and RL algorithms. The aim was to minimise the rate of paw-to-surface contact and thereby reduce the loudness of the footsteps. The system uses data from sensors on the paws and imposes "penalties" for movements that are too abrupt.

1. Experiment: The researchers compared the new technique with RL-based controllers as well as the official Sony controller.
2. Result: The new model was able to significantly reduce noise - the aibo walked noticeably quieter than with Sony's proprietary algorithm and other standard RL approaches.

Dancing instead of simple movements

The second part of the team's work focuses on how to teach aibo to effectively "dance" and interact with others. To do this, they developed the Deep Fourier Mimic (DFM) model. This technology combines motion representation and reinforcement learning, allowing the robot to mimic given dance patterns and supplement them with additional tasks, such as moving.

• Themain "trick": Instead of a simple "playback" animation, the robot is able to move naturally between different movements and react to the user's actions.
• Benefit: Smoother and more "alive" manoeuvres: aibo can visually "dance" to a human and looks much more synchronised.

Prospects and limitations

The researchers believe that the new RL algorithms can be integrated not only in aibo, but also in other home or entertainment robots (e.g., amusement parks). However, some subtleties were observed during testing:

1. Balance between quiet walking and stability: The slower the contact speed of the paws, the higher the risk of losing stability; future models should take into account the environmental conditions (floor surface, presence of obstacles).
2. Limited usefulness for non-periodic movements: DFM works well for rhythmic tasks (stepping, dancing), but is not as effective for actions such as grasping or lifting from the floor, which for now need to be trained separately.

But even with these nuances taken into account, the work paves the way for aibo to become an even friendlier "companion animal",