GO-2

GO-2 (Genie Operator-2)

AgiBot · 2026-04(2차)

VLANew in 2026Not disclosedNot disclosed

Combines Action CoT, which plans directly in action space, with an asynchronous dual system.

Paper

Backbone LLM: ViLLA framework (secondary)
Vision encoder: —
Action head: Action CoT + asynchronous dual system (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): ViLLA framework (secondary); Action Chain-of-Thought (macro planning in action space) + asynchronous dual system (low-frequency semantic planning System 2, high-frequency action tracking System 1, residual refinement)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed; distributed training across thousands of robots with Genie Studio, about 10× training efficiency·2–4× success·50%+ data reduction (company claim)
Data type: Interaction data (composition not disclosed)
Data scale: Tens of thousands of hours
Main datasets: Not disclosed
Tasks & evaluation: LIBERO average 98.5%; LIBERO-Plus zero-shot 86.6%; VLABench average 47.4; Genie Sim 3.0 sim→real 82.9%
Robot: AgiBot robot (model not stated)

Reinforcement learning: Not disclosed · Training method not disclosed
Base control: Fixed · No mention of chassis·locomotion control (target platform not stated)

Key features: Combines Action CoT, which plans directly in action space, with an asynchronous dual system. Trained only in simulation, it reported 82.9% real-world success.
Notes: Corrections applied: the only source with a verifiable date is TRR 2026-04-09 (secondary, 'today introduced'). AgiBot page 2026-09-08† appears only in an image path, so unverifiable. Action CoT accepted at CVPR 2026, dual system at ACL 2026.
Verification: Corrections applied