GE-Act 2.0

GE-Act 2.0 (Genie Envisioner Act 2.0) is a world-action model for robotic manipulation presented by AgiBot Research. World-action models predict future states to guide robot actions, enabling learning from both action-free video and action-labeled interaction. The authors state that most existing world-action models inherit pretrained video generators, leaving pretraining and scaling of such models underexplored, which is the problem GE-Act 2.0 addresses. Architecture: GE-Act 2.0 combines a control-oriented autoencoder (CoAE), a single-step visual planner (SVP), and an inverse dynamics model (IDM). Every trainable generative and action component is initialized from scratch on manipulation data. CoAE retains action- and instruction-relevant information under aggressive compression. SVP produces a completed future state in one differentiable pass, allowing visual planning and inverse dynamics to be pretrained separately on complementary data. The components are then jointly trained using knowledge-aligned selective optimization (KASO), which reduces mismatched supervision by selecting only predicted future states judged behaviorally compatible with the recorded action. Reported results, as stated by the authors: pretrained checkpoints were evaluated directly, without per-task fine-tuning, on 100 tasks across 20 manipulation skill groups with held-out scenes, backgrounds, lighting and object instances. Scaling co-training data from 300 to 30,000 hours raised success from 17.1% to 44.1% on G1-OP and from 13.4% to 31.1% on G2-90D. G2-90D comprises less than 2% of the co-training data yet improved by 17.7 points, which the authors suggest indicates cross-embodiment transfer. Gains span 19/20 and 18/20 skill groups, and skill-specific coverage correlates with zero-shot out-of-distribution success (Pearson r=0.80; Spearman rho=0.85). Under the same protocol, the model grounds object, color, shape and position references in at least 90% of trials. The paper is authored by the AgiBot Research Team and is accompanied by an official project page at ge-act-v2.github.io.