The first step was to design the virtual training environment for our algorithms in a physics simulation software (Bullet Physics). We chose parameters (e.g., gravity or the motor impulses in the robot) to recreate realistic circumstances. We constructed the initial basic kinematic kit for our simulated robot, Babu.

Slide right to see more. >>

We started with very simple models and moved towards more complex ones. We’ve also built a developer interface for the supervised part of the learning process. We were implementing the very first algorithms so that our robot could practice and improve its previously learned skills in an unsupervised fashion.

To get to the first steps of Babu, we closely followed the development path of a human baby learning to walk on two legs after becoming aware of its limits and learning to balance its core, muscles, and movements. Once the core skills were sufficiently developed, the unified and hierarchical system allowed Babu to build on it autonomically.

This process not only improves efficiency but also finds a large amount of style variations that all retain bipedal balance. In the current stage of development Babu is already able to adapt his skills to new circumstances such as a different constitution, or special tasks like walking in high heels.

Babu can manipulate objects of different weights. In this video, the luggage weighs around one-third of the weight of Babu. Note that Babu does not need to stop and get into a 'lifting' position when picking up the luggage. The AI takes these into account when establishing bipedal balance.

We implemented several training rooms for Babu to practice his skills under variable circumstances. These setups often require him to react to random events that are unpredictable for the AI. These “training sessions” serve to improve the adaptivity of Babu’s kinematic kit. Of course, there is no learning without failing.

The high degree of adaptivity reached during these sessions is a critical component of making Babu interact with users. It means that Babu becomes able to follow real-time kinematic instructions coming from a VR or motion capture device while retaining the key elements of his own set of moves that are necessary to maintain his balance.

The following scenes demonstrate real-time interaction with Babu. The resulting motion always manifests as a combination of previously obtained skills and an intent to adjust to user input. In this video, Babu has already learned how to move his weight from one foot to another and execute a periodic base stance that behaves well under random permutations. This stability is essential for achieving steady control during the interaction.

The main thing that sets apart our design from usual computer gaming control is that one cannot directly force its will on what is happening on the screen, since it must be consistent with both the laws of the simulated physics and Babu's existing kinematic kit. The teaching of Babu therefore begins with a learning process for the user itself, whereby the user has to explore how the system reacts to a variety of inputs.

Input at this stage is captured with its full dynamics, and Babu can sensitively react to small changes in velocity or position. When control is executed in good synchronization with Babu's motion, the user and the AI can form new moves in cooperation, leading back to another stable periodic state. Collaboration requires learning from both the human and the AI.

After reaching a suitable degree of stability, we can turn the external control of Babu into internal control. In other words, Babu records the moves shown to him into his kinematic kit and can reinforce them via unsupervised training. Symbolic commands can later invoke these moves without showing the exact dynamics again.

Improvements on the base kinematic kit allow us to teach increasingly complex motions to Babu, including manipulating objects during his exercises or adjusting to challenging environments during the interaction. Eventually, we even managed to teach Babu to do a proper handstand.

We created a MaxWhere showroom to showcase Babu's abilities. MaxWhere is a universally available and easy-to-use 3D virtual platform with a captivating visual world, proven to support the reception and comprehension of virtually shared information and improve collaboration. Collaboration is king.

Our proof-of-concept game demonstrates the capabilities of the technology. Humans and AI need to work together to achieve a common goal. There was a lot of background work to ensure the game was accessible from a web browser. We can simultaneously simulate up to four robots, enabling the user to have a friendly multiplayer tournament.