Event

AgiBot Releases GE-Act 2.0 World-Action Model

1 reports · 1 sources

The AgiBot Research Team released a technical report on the GE-Act 2.0 world-action model. All trainable generative and action components of the model were trained from scratch on manipulation data, and it consists of a control-oriented autoencoder (CoAE), a single-step visual planner (SVP), and an inverse dynamics model (IDM), jointly trained with KASO. The team reports that it directly evaluated the pretrained checkpoint on 100 tasks across 20 manipulation skill groups without per-task fine-tuning, with test scenes, backgrounds, lighting, and object instances all held out. After scaling the joint training data from 300 hours to 30,000 hours, the G1-OP success rate rose from 17.1% to 44.1%, and G2-90D rose from 13.4% to 31.1%; G2-90D accounted for less than 2% of the data yet improved by 17.7 percentage points, which the team takes as evidence of cross-embodiment transfer.