Data availability, data reliability, and the challenge of objective evaluation are the three key factors limiting AI adoption in animation and robotics, according to ML/AI engineering expert and founding ML engineer at Uthana, Dave Savostyanov. In an exclusive interview with The AI Journal, Savostyanov discusses the tools he has developed to overcome these challenges, unlock the full potential of generative AI, and how these approaches are gradually becoming an industry standard.
On the future of AI in computer animation and robotics
Q1. How widely is AI being adopted in these fields today, and to what extent is its potential being realized?
“AI is already used extensively across the animation field, from concept development to image enhancement, facial animation, motion generation, and more. Robotics is a more challenging field because the cost of failure is significantly higher. If a physical system behaves incorrectly, it can damage its surroundings or even injure someone, whereas in animation, the worst-case scenario is simply a poor-quality result. That said, AI is also being actively adopted in robotics. Most applications are still at the research or prototype stage, although there are already several highly successful examples that have moved into large-scale deployment, such as self-driving cars.”
From data scarcity to the challenges of evaluation
Q2. In your view, what factors are slowing this process down?
“I would highlight three key factors. The first is data availability. In animation, collecting data requires hiring actors, renting a studio and motion capture suits, recording the sessions, and processing the results. It’s a slow and resource-intensive process. A recording session typically lasts no more than 12 hours, whereas training a model ideally requires hundreds of thousands or even millions of examples. On top of that, errors are inevitable. An actor may accidentally knock a motion capture marker out of place, and suddenly the reconstructed 3D character is floating above the ground instead of walking on it. All of these issues have to be detected and corrected.
The second factor is reliability. In animation, a character floating above the ground simply results in a poor visual output. In robotics, however, the same kind of error can cause a robot to stumble, fall, misjudge its surroundings, or behave unpredictably.
The third challenge is quality evaluation. Humans can usually look at a 3D scene and immediately recognize that something is wrong. Translating those observations into objective numerical metrics that a machine can understand is far more difficult. Without reliable metrics, it’s extremely challenging to improve model quality or meaningfully compare the results of different experiments.”
Model pretraining and the power of collective intelligence
Q3. What proprietary methods and tools have you developed to address these challenges?
“When it comes to quality evaluation, we’ve developed a new approach by involving as many people as possible in assessing model outputs. One way we do this is through targeted advertising. For example, someone using a language-learning app might see one of our generated videos between lessons and be asked to rate how well the animated motion matches the text prompt. Of course, we rely on a comprehensive set of internal evaluation metrics, but incorporating this form of collective human feedback has proven to be extremely valuable.
On the data side, we’re actively exploring model pretraining through motion capture from video. This approach allows us not only to increase the volume of training data by orders of magnitude, but also to make it far more diverse and representative. Pretraining itself is not a new concept. Until recently, however, it was used primarily for large language models such as ChatGPT, Claude, and Gemini. We were among the first to apply the same principles to 3D generative models. Today, we’re seeing other companies and studios adopt this approach as well, and I believe it has the potential to become a new industry standard.”
From hypothesis to solution: the advantages of a scientific and engineering approach
Q5. In a previous interview, you mentioned that you rely on a scientific and engineering approach of your own design when solving applied problems. What does this approach involve, and what are its advantages?
“Our approach is to treat every engineering challenge as a scientific research problem. We formulate hypotheses, design experiments that can be quantitatively validated, and use the results to derive objective performance metrics. I see this methodology gaining increasing traction across the industry, with leading companies adopting it as a core part of their development process.”
Contributing to the industry through conferences and professional communities
Q6. You are a member of leading industry associations, regularly attend conferences, and give talks and lectures. Why do you believe these activities are important, and how do they contribute to the development of the industry as a whole?
“Historically, the Tech industry has been shaped by a strong open-source culture. When companies publish their research and technical work, and when their experts share knowledge through conferences, lectures, and professional associations, the entire community benefits. It accelerates innovation across the industry. That’s something I strongly believe in, and I make a conscious effort to contribute whenever I can.”
Q7. You were recently recognized as an expert by the prestigious Hackathon Raptors association. Could you tell us more about that achievement?
“My involvement with hackathons began back when I was a university student. I was fascinated by the way they combine competition, innovation, and teamwork. Over time, I transitioned from being a participant to serving as a judge and organizer, a role I’ve held for more than a decade.
Judging hackathons exposes you to an incredible range of ideas, concepts, and experimental solutions, many of them highly creative and completely unexpected. I believe that my long-term commitment to the community and active involvement over the years ultimately led to this recognition.”
