The open‑source revolution in robotics: four forces and the games behind the "free brain"

Open brains, big bets

A wave of companies — from Xiaomi (小米), Ant Group (蚂蚁集团) and Alibaba Damo Academy (阿里达摩院) in China to NVIDIA (英伟达) and Google (谷歌) globally — have begun openly publishing robot “brains”: models, weights and training toolchains. They are giving away core software for embodied intelligence at a time when robots are moving from lab curiosities to commercial platforms. Why hand away your crown jewels? Because open source can seed ecosystems, lock in platforms, and attract partners for hardware, cloud and simulation services — all while competing for a trillion‑dollar future.

Architectures and contests

The technical contest centers on variants of VLA (Vision‑Language‑Action) models. It has been reported that OpenVLA — a compact, 7‑billion‑parameter open model using two visual encoders (DINOv2 for spatial structure and SigLIP for semantics) plus Llama 2 for fusion — beat Google DeepMind’s much larger RT‑2‑X on a 29‑task benchmark, reportedly outperforming the 55‑billion‑parameter RT‑2‑X by 16.5%. Smaller, more flexible projects also matter: “Octo” targets broad generalization with only tens of millions of parameters to enable zero‑shot transfer across robot platforms. It has been reported that NVIDIA’s GR00T N1 series — promoted as an “open humanoid base model” — couples a slow, deliberative System 2 (vision‑language planner) with a fast, action‑level System 1 and bundles simulation and data tools (Omniverse, Isaac Sim, Newton) to sell an end‑to‑end stack.

Players and Chinese momentum

China’s contributors are expanding beyond followership into rule‑setting. Xiaomi‑Robotics‑0 (Xiaomi, 小米) and Ant Group’s LingBot‑VLA (蚂蚁集团) are examples: Xiaomi’s 4.7‑billion‑parameter MoT hybrid separates “big brain” planning from “small brain” execution to cut latency, while LingBot pretrains on tens of thousands of real‑robot hours to pursue cross‑morphology control. Academic teams — Tsinghua AIR (清华AIR) and Shanghai AI Lab (上海AI实验室) with X‑VLA — have open‑sourced code, data and weights as rigorous baselines. Startups such as Xinghaitu (星海图), Zhiyuan Robotics (智元机器人) and Xingdong Epoch (星动纪元) are opening datasets or shipping models to real robots. Meanwhile Google has swung between openness and closure — RT‑1 was open, later RT models closed — and it has been reported that Google and Boston Dynamics are deepening collaboration around Gemini Robotics to make a common robotics “Android.”

Geopolitics, risks and the next move

This open‑source rush runs against a tricky geopolitical backdrop: export controls on advanced chips and rising U.S.‑China tech tensions. Open models lower the barrier to entry for research and startups inside and outside China, but they also become strategic assets — a way to cultivate customers for simulation, cloud GPU time or proprietary sensors. So is the “free brain” a democratizing force or the newest vehicle for platform lock‑in? Expect more open releases, rapid forks and commercial plays as companies race to turn community momentum into durable advantage.