Asmov-1's team reports a 0.007% error rate across all evaluation tracks and expects an open-source release in early November.
In an October 31 update, the team said Asmov-1 training uses Cartwheel cross-embodiment, control and its scaling laws dataset. It expected to release the policy as open source the following week alongside RoboParty RP1, whose training it said was finished.
3.1K
2 Sources, first seen ago
@andrew_n_carrAsmov-1 policy is 85% of the way done training using Cartwheel cross-embodiment, control, and our scaling laws dataset. We're sitting at 0.007% error rate on all eval tracks. Open source release should be next week (RoboParty RP1 policy finished training and will be included too).
@ericjang11@andrew_n_carr So cool! 🦾@andrew_n_carrAsmov-1 policy is 85% of the way done training using Cartwheel cross-embodiment, control, and our scaling laws dataset. We're sitting at 0.007% error rate on all eval tracks. Open source release should be next week (RoboParty RP1 policy finished training and will be included too).@ericjang11@andrew_n_carr So cool! 🦾@andrew_n_carrAsmov-1 policy is 85% of the way done training using Cartwheel cross-embodiment, control, and our scaling laws dataset. We're sitting at 0.007% error rate on all eval tracks. Open source release should be next week (RoboParty RP1 policy finished training and will be included too).@ericjang11@andrew_n_carr So cool! 🦾@andrew_n_carrAsmov-1 policy is 85% of the way done training using Cartwheel cross-embodiment, control, and our scaling laws dataset. We're sitting at 0.007% error rate on all eval tracks. Open source release should be next week (RoboParty RP1 policy finished training and will be included too).@ericjang11@andrew_n_carr So cool! 🦾—
Not ranked yet
—
Not ranked yet
