Who's winning.

Known biases in this draft

• Disclosure bias. Only 32 of 102 tracked OEMs disclose a funding figure. A well-funded company that stays quiet scores lower than a louder, smaller one. This is the biggest flaw.
• Public and internal programs are penalised. Tesla and Honda fund from the balance sheet, so they score zero on capital despite obvious resources.
• Capital is log-scaled from a $1M floor. An earlier version measured from zero, which squeezed every real funding figure into the top of the scale ($5M scored 0.71 against $2.34B at 0.99) and pushed companies that raised a tenth as much above Figure AI. Measuring from a floor spreads the range properly, but it is still a log: a company that raised 10x more does not score 10x higher, by design.
• Market capitalisation is ignored. Capital counts money raised, not what the company is worth. Unitree raised ~$905M in its IPO and trades near $66B; only the raise is scored, so listed companies read smaller here than they are.
• Deployment is counted from free text. Partners are read out of a prose field, so a named site can read as a second partner. Prose that names nobody now scores zero rather than counting as partners, but this stays a coarse 0–4 proxy, not a customer count.
• No capability measure. Nothing here scores what the robots can actually do. That needs benchmark data the dataset doesn’t have yet.
• Sectors are scored independently. A score of 60 in prosthetics and a 60 in humanoids reflect standing within very different, much smaller or larger peer sets, not the same thing.