It is not just limited to humanoids, as we are training better robotics "base" model.
Currently we are testing it on randomized LIBERO-PRO benchmark, in stage II we will launch our own benchmark and introduce real robot deployment as gate.
Finally we will work with factories & humanoid companies to deploy specific robotics policy for specific working envs.
Introducing OpenRoboto, an open competition on Bittensor for continuously improving robotics models. We acquired subnet 80 and now is fully live on mainnet.
How it works: pull the current base model (we start from π0.5), fine-tune it on the openpi runner, upload the weights to Hugging Face, and announce the exact commit on-chain. Every round, challengers are scored on randomized LIBERO-Pro Benchmark; only a strictly-better model becomes the new base and earns the emission.