AgiBot ships GE-Act 2.0 with claimed zero-shot scaling on embodied tasks
The AMW Read
Explicit 100x data-scaling metrics on a from-scratch embodied world-action model meaningfully update a tracked robotics player’s baseline without resolving the wider embodied-foundation debate.
AgiBot ships GE-Act 2.0 with claimed zero-shot scaling on embodied tasks
AgiBot released GE-Act 2.0, a native world-action model trained from random initialization on embodied manipulation data rather than fine-tuned for evaluation tasks. Per company figures carried by PingWest, training data expanded from 300 hours to 30,000 hours. On a 100-task atomic suite, G1-OP success is reported to have risen from 17.1% to 44.1%, with solvable tasks climbing from 39 to 76, including finer skills such as folding towels and nesting paper cups. The stack uses CoAE visual representation, an SVP single-step visual planner, and KASO joint training; inference is cited at 104 milliseconds on a single RTX 5090, and the firm says G2-90D body data was under 2% of the mix while still claiming capability transfer.
Physical AI has been split between task-specific demos and the bet that foundation-style scaling can generalize in the real world. A from-scratch train paired with a zero-shot, no-eval-tuning protocol and a roughly 100x data step is a measurable claim in that contest: success moving with data volume, not with narrow benchmark fitting. Per the AI Market Watch index, AgiBot-matched pipeline items rose to 4 in the last 90 days from 1 in the prior window (name-matched over ingested sources only), consistent with a denser release cadence around this player.
Builders and investors should treat the scaling narrative as a protocol to pressure-test—zero-shot on unfamiliar scenes, no task-specific fine-tuning—rather than as settled law. The near-term tell is whether peers can reproduce similar 300-to-30,000-hour curves and atomic-task lifts on comparable suites, and whether cross-body transfer with thin embodiment-specific data holds outside AgiBot’s own robots. If those results replicate, capital will keep concentrating on manipulation data factories and evaluation harnesses, not just polished single-body demos.
