Remi Fabre drops the third episode of his live-action robot series — and the brain behind his duck

The duck that stole the show

Fabre, a robotics researcher focused on natural interaction, posted the third installment of a series shot entirely with physical robots. No animation. No editing that hides the glitches. The first take was a disaster; he nearly called off the shoot. Then he reviewed the raw footage and realized the Microduck, the tiny robot at center stage, had delivered what he calls an "Oscar-level performance" precisely because everything went sideways. The most convincing moments emerged from the chaos — a reminder of the gap between polished demos and real-world behavior.

An open-source emotion dataset

Alongside the video, Fabre released a dataset on Hugging Face containing the "emotions" authored for the sketch: a library of behavioral states that let the duck express curiosity, frustration, joy, and so on. The data is free to use, with an explicit invitation not just to consume but to create new states and contribute them back. The approach echoes motion libraries in robotics, except the primitives are expressive behavior patterns rather than walking trajectories, and the license permits commercial embedding without special restrictions.

Quackd: a natural-language brain for the duck

The bigger announcement is quackd, a planning layer that sits atop the robot's existing skills. A user writes a free-form goal — "find the ball and kick it" — and a large language model decomposes it into a sequence of actions the duck already knows: perceive, navigate, manipulate. The system also supports local models, so there's no cloud dependency or API key requirement. Goals live in `.duck` files, a plain-text format analogous to declarative config: define *what*, not *how*.

Modular architecture over monolith

Separating the planning layer (quackd) from the execution layer (the existing skills) means you can swap the language model without rewriting motor control, and vice versa. It's a familiar pattern in modern robotics — task planning divorced from motion planning — but the implementation here is deliberately minimal: one text file per goal, a closed skill set, and an LLM broker translating between them. The upside is transparency and debuggability; the downside is the system is bounded by what the skills cover, with no ability to learn new motions from the prompt.

What's missing from the picture

Fabre published no benchmarks — no success rates, latency figures, or comparisons against classical planners or end-to-end vision-language-action models. It's also unclear which LLM served as the default in the demo, or how the system handles a single skill failure mid-sequence. The Hugging Face dataset is live, the quackd code is open, and anyone with a compatible duck (or similar platform) can run it themselves. It's a solid starting point, not a finished product.