In the first half of 2026, more than 93.5 billion yuan poured into embodied AI across 322 funding rounds, pushing the sector to a fever pitch. Arriving with the capital was another high-frequency word: ‘world number one’. By one count, over 20 leaderboards are now active, each minting a new ‘first’ almost weekly.
The speed of ‘first’ now outruns the speed of robot iteration. A model can go from launch to chart-topper in a week, and the charts refresh faster than products ship. When ‘first’ multiplies, the industry slides into a collective confusion about ranking inflation, and a fair question: who is actually manufacturing these titles?
A manufactured supply chain
For a startup, a ‘#1’ is never a medal, it is a commercial tool. Embodied AI is still a story-driven track with no settled technical route and no proven business model, so investors lack hard metrics and reach for rankings as the most intuitive valuation anchor. One title can lift a valuation by hundreds of millions and unlock the next round. Qianxun and Star Era both announced funding right after topping boards, an open industry routine: build hype with a ranking, lift the valuation with the hype, close the round.
Where demand is strong, supply appears. Today’s embodied-AI leaderboards fall into three families: operation benchmarks for real manipulation, world-model benchmarks for physical reasoning, and specialised benchmarks for edge cases. The problem is that some are commercial products with opaque, negotiable rules tied to business cooperation. At the June Zhiyuan conference, institute head Wang Zhongyuan said plainly that rankings are ‘not fully trustworthy’ and too numerous to parse. The RoboArena case exposed a distributed-eval flaw: anyone can register as an evaluator, so candidates can grade themselves.
Closed evals have their own loophole: overfitting. In a 2026 hackathon, a university team tuned its model to a fixed fruit-sorting task on the public leaderboard, then collapsed on the blind test with new fruits and a reshuffled table. That is ‘exam-oriented education’ for robots, and it is the opposite of the generalisation embodied AI is supposed to deliver.
No ruler yet
Step back and the root cause is simple: the whole sector is running blind without a shared reference frame. Cars have 0-100 acceleration, range and ADAS tiers, uniform and reproducible. Phones have chip processes and benchmarks. Even the newer LLM field converged on MMLU, GSM8K and HumanEval. Embodied AI has not yet built that consensus, and until it does, the ranking carnival will keep drowning out the technology underneath.
Read the original report (OFweek Robotics)
Translated and adapted from OFweek Robotics (robot.ofweek.com).