AgiBot has released GE-Act 2.0—a native, world-class model trained exclusively on embodied manipulation data, rather than being adapted from an existing video generator. The company asserts that this release represents the first systematic investigation into the possibility of a world-class model unlocking new skills for real robots as training hours increase by two orders of magnitude—from approximately 300 hours to 30,000 hours—without requiring fine-tuning for a specific task or using demonstration clips during evaluation.
Unlike existing systems that attach an action head to a frozen video model, GE-Act 2.0 simultaneously trains visual representation, future prediction, and action prediction on robot data. During scene testing, backgrounds, lighting, and object instances are excluded from the training process. Within this zero-shot protocol, covering 100 atomic tasks and about 20 skill families across two embodiments, success rates increased with the volume of data: on the G1-OP platform, overall success rose from 17.1% to 44.1%, and on G2-90D—from 13.4% to 31.1%, with the gain remaining noticeable between the 5,000 and 30,000-hour levels.
The company highlights the 'ignition' phenomenon for fine skills. Tasks showing non-zero success expanded from 39 to 76 on G1-OP and from 24 to 72 on G2-90D. Actions such as folding a towel, placing paper cups, removing and replacing pen caps, and arranging a bouquet in novel scenes only began to reliably manifest after reaching the largest data level. Cross-embodiment transfer was also observed: G1-OP provided over half of the training data, while G2-90D accounted for less than 2%, yet the rarer body still gained about 17.7 percentage points as total data grew.
Architecturally, the CoAE visual encoder compresses each 256x384 frame into 24 tokens while preserving semantic, motion, and local structural cues. A one-step visual planner predicts a full future segment, and a reverse dynamics model translates the predicted change into actions; AgiBot notes that this takes about 104 milliseconds per action chunk on an RTX 5090. The Knowledge-Aligned Selective Optimization method selects several future scenarios and retains the one that best matches the recorded actions to reduce the discrepancy between planning and action.
The data recipe involves mixing incomplete streams: hours of video instructions for pre-training the world model (including egocentric frames without action labels), trajectory hours for reverse dynamics, which include failures and rollouts, and fully paired hours of video instructions and actions for joint training. In AgiBot's GE stack, GE-Sim is aimed at policy evaluation and environment construction, while GE-Act translates predicted future states into robot movement. After optional fine-tuning, AgiBot reports a score of 60.52% in out-of-distribution RoboTwin tests and 0.770 in GenieSim instruction following, surpassing several comparable baseline models. Project materials are available at ge-act-v2.github.io.