Hugging Face's $399 Microduck: an open-source robot you train with RL
Apache-2.0 training stack, sim-to-real in ~2 GPU-hours, and a beak that grabs socks. Plus: htmx 4.0 goes fetch()-native and a new science-agent benchmark.
Copy markdown
The $399 duck that learns like a model
Pollen Robotics (a Hugging Face company) opened pre-orders on Microduck, a 25cm, sub-800g biped with 15 motors, a Rockchip RK3566 + AI accelerator, a camera, mics and an 8x8 LiDAR. It ships seven behaviors out of the box — walk, sit/stand, kick, grab, roller-skate and self-recovery — running a 50Hz policy loop.
The whole training stack is Apache-2.0
The RL pipeline is open on GitHub (microduck_rl): train new behaviors in simulation with mjlab (MuJoCo Warp) + PPO across 4,096 parallel environments in ~1–2 hours on a single CUDA GPU, then deploy sim-to-real via a BAM M6 actuator model with domain randomization. It's the cheapest end-to-end 'RL on real hardware' loop yet — a genuine weekend project.
The catch: closed hardware, delivery by Christmas
Only the software is open — the mechanical and electronic design files stay proprietary, so you can teach it tricks but can't fork the body. Runtime is roughly one hour on a removable NP-F550 battery, and units are slated to ship before Christmas 2026.
htmx 4.0 'The Fetchening' ships
htmx cut its first major release since 1.0, rebuilt around the browser's native fetch() API. It's a breaking change with a new default request model, so pin your version before touching production — but greenfield projects get a cleaner, more modern core.
A benchmark for science agents
Terminal-Bench-Science 0.1 launched to grade coding agents on real research workflows across scientific domains, and it's open to community task contributions — so you can submit the exact workflow your agent keeps failing and see how the frontier models score.