About
GO-1 (Genie Operator-1) is a family of open VLA foundation models for robotics developed by OpenDriveLab (HKU + Shanghai Innovation Institute) and AgiBot. All family models are pretrained on the AgiBot World dataset (Alpha 92k trajectories + Beta 1M trajectories), use InternVL 2.5-2B as the base Vision-Language Model, and are publicly available on HuggingFace under CC BY-NC-SA 4.0. The family includes: (1) GO-1 — full model with Latent Planner and Action Expert; (2) GO-1 Air — version without Latent Planner, faster and lighter (3B parameters); (3) GO-1-Pro — next iteration with a velocity model filtering velocity multimodality. Models pretrained on a single embodiment (AgiBot G1) exhibit cross-embodiment transfer to other platforms (AgileX, Franka, RoboTwin) with fine-tuning on <200 demonstrations. A reference point for many academic works — IROS 2025 Best Paper Award Finalist & IEEE TRO 2026 paper.
