Q&A with Aaron Saunders, VP at Boston Dynamics, on teaching robots to dance and how that informs the company's approach to robotics for commercial applications
Evan Ackerman / IEEE Spectrum : Tweets: @pranavab47 , @aparanjape , @mouserelec , @kaeshour , and @helengreiner Tweets: Pranav Bhounsule / @pranavab47 : Boston dynamics does not use machine learning to control their robots. https://twitter.com/... Amit Paranjape / @aparanjape : By now, most of you would have seen those amazing ‘Dancing Robots’ videos by Boston Dynamics (over 24 million YouTube views in a week). Here's how they did it... quite interesting: https://spectrum.ieee.org/... Mouser Electronics / @mouserelec : Wondering how those @BostonDynamics robots got their sweet dance moves? 💃 How Boston Dynamics Taught Its Robots to Dance: https://mou.sr/... Image: Boston Dynamics https://twitter.com/... @kaeshour : Excellent piece: 1 knowledge = heuristic error-correction, 2 substrate-independence of human creativity: thought-music-motion-code 3 how human-driven machine learning is (dancers scripted moves to music; engineers abstracted them to machines' repertoire) https://spectrum.ieee.org/... Helen Greiner / @helengreiner : 24 Million Views!...that beats “cat on roomba”. Your move @irobot.... How Boston Dynamics Taught Its Robots to Dance - IEEE Spectrum @ieee @BotJunkie @BostonDynamics @ieee @ieeespectrum https://spectrum.ieee.org/...
Context & Ripple Effects
By January 2021 Boston Dynamics had already crossed from lab curiosity to product: after years of struggling to articulate what its robots were actually for, it was shipping Spot commercially while managing the public-perception problem that came with viral videos. The dance clip — over 24 million YouTube views in a week per the related tweets — is the marketing side of that tension, and VP Aaron Saunders uses the Q&A to explain the engineering behind it.
The notable technical claim circulating alongside the piece, echoed by roboticist Pranav Bhounsule, is that Boston Dynamics does not use machine learning to control its robots — the choreography comes from model-based control. That stance is what makes this interview a useful marker: two years later Saunders was back answering questions about generative AI's role in robotics, and by 2025 the company had built a large behavior model with Toyota Research Institute producing emergent skills in humanoids.
First-order effects
- Saunders' explanation positions Boston Dynamics' whole-stack, model-based approach as a deliberate alternative to the learning-based methods spreading through robotics research — a positioning the company must defend as customers evaluate Spot for real work sites.
- The viral reach of the dance video gives Boston Dynamics cheap brand awareness at exactly the moment its first commercial robot needs buyers, converting entertainment attention into sales-funnel interest.
Second-order effects
- Rivals and research labs leaning on reinforcement learning face a high bar set by polished demonstrations of hand-engineered control, pressuring them to match the showmanship as well as the capability.
- As generative AI matures, the no-ML doctrine becomes untenable — the later DeepMind Gemini integration into Atlas and the Toyota behavior-model work are the visible consequences of that pressure.
Third-order effects
- If the pattern holds, humanoid robotics consolidates around foundation models for movement and manipulation, with companies like Boston Dynamics shifting from hand-tuned controllers to learned behavior stacks — a reversal of the philosophy this interview articulates.
- Viral demonstrations become a standing requirement for commercial robotics firms: public perception, flagged as a core challenge before Spot's launch, effectively functions as a marketing channel that funds and legitimizes hardware programs.
The trend: Robotics is moving from hand-engineered, model-based control toward AI-learned behavior models, with Boston Dynamics' own trajectory from dance choreography to foundation-model humanoids tracing the industry's pivot.