AgiBot has released GE-Act 2.0, a world-action model trained from random initialization on embodied manipulation data, providing what the company calls the first systematic test of whether scaling robot training data by two orders of magnitude — from roughly 300 hours to 30,000 hours — can unlock new real-robot skills without task-specific fine-tuning or demonstration clips at evaluation time1.
Unlike pipelines that graft an action head onto a frozen video generator, GE-Act 2.0 trains vision representation, future prediction, and action prediction jointly on robot data.
Zero-shot results across two embodiments
AgiBot evaluated the model on a zero-shot protocol covering 100 atomic tasks and roughly 20 skill families across two hardware platforms, with scenes, backgrounds, lighting, and object instances held out of training. On the G1-OP platform, overall success climbed from 17.1% to 44.1% as data scaled; on G2-90D, it rose from 13.4% to 31.1%, with gains still visible between the 5,000-hour and 30,000-hour tiers.
The company highlights a "light-up" pattern for fine manipulation skills: tasks showing non-zero success expanded from 39 to 76 on G1-OP and from 24 to 72 on G2-90D. Behaviors such as towel folding, nesting paper cups, removing and replacing pen caps, and flower arranging in novel scenes appeared reliably only once the largest data tier was reached.
Cross-embodiment transfer also emerged. G1-OP contributed more than half of training data while G2-90D accounted for under 2%, yet the scarcer body gained about 17.7 percentage points as shared data grew.
Architecture and inference
The model uses a CoAE visual encoder that compresses each 256-by-384 frame into 24 tokens while retaining semantic, motion, and local structure cues. A one-step visual planner predicts a full future chunk, and an inverse dynamics model maps that predicted change into actions. AgiBot cites about 104 milliseconds per action chunk on an RTX 5090.
A technique called Knowledge-Aligned Selective Optimization samples multiple futures and keeps the one most consistent with recorded actions, reducing planning-action mismatch. The data recipe mixes instruction-video hours for world-model pretraining — including egocentric footage without action labels — trajectory hours that include failures and rollouts, and fully paired instruction-video-action hours for joint training.
After optional fine-tuning, AgiBot reports 60.52% on RoboTwin out-of-distribution tests and 0.770 on GenieSim instruction following.
Within AgiBot's broader GE stack, GE-Sim targets policy evaluation and environment building, while GE-Act turns predicted futures into robot motion. Project materials are posted at ge-act-v2.github.io.
ANALYSIS The results bear directly on a central question in embodied AI: whether data scaling laws analogous to those observed in language models hold for physical manipulation. The persistent gains between the 5,000-hour and 30,000-hour tiers, and the emergence of fine skills only at the largest tier, suggest the scaling curve has not yet saturated for this architecture.