GO-1

GO-1 (Genie Operator-1)

AgiBot·HKU·Shanghai Innovation Institute·Shanghai AI Lab · 2025-03

VLAIndustry models3Bdiffusion

Plans with embodiment-agnostic latent actions learned from human video, and an Action Expert converts them to real actions.

Paper

Backbone LLM: InternLM2.5-1.8B · InternVL2.5-2B config
Vision encoder: InternViT-300M
Action head: Latent Planner (24 layers) → diffusion Action Expert (diffusion)
Chunk: 30 steps
Control rate: Not disclosed
Architecture (original summary): ViLLA: InternVL2.5-2B VLM + Latent Planner (24 layers) + diffusion Action Expert; H=30 action chunk; latent actions learned from human video with VQ-VAE-based LAM

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: GPU and time not disclosed; 30,000 steps, batch 768, lr 2e-5; README memory: full ~70GB (batch 16, A100 80GB · H100), Action Expert only ~24GB (RTX 4090 · A100 40GB)
Data type: Real-robot teleoperation + human video
Data scale: AgiBot World 1,001,552 traj (paper)/1,003,672 traj (GitHub Beta), 2,976.4 h, 217 tasks, 100 robots, 106 scenes, 87 skills
Main datasets: AgiBot World (Alpha 92,214 traj, Beta), Ego4D (latent action learning)
Tasks & evaluation: AgiBot G1 real-robot tasks (Restock Bag, Table Bussing, Pour Water, Restock Beverage, Fold Shorts, Wipe Table; 30 trials per task = 10 seen + 20 variants); +32% over RDT; AgiBot World pretraining improves average by 30% over OXE
Robot: AgiBot G1 (7-DoF bimanual, mobile chassis, adjustable waist; includes 6-DoF dexterous hand data)

Inference device & speed: README: about 7GB for inference, RTX 4090; latency not disclosed
Inference memory: About 7 GB (RTX 4090, README)
Weights bf16 (calc.): 6 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Latent action pretraining + imitation learning
Base control: Fixed · No mention of base (chassis)·waist actions (evaluation tasks are tabletop manipulation; chassis mentioned only in teleoperation collection)

Key features: Plans with embodiment-agnostic latent actions learned from human video, and an Action Expert converts them to real actions. Trained on AgiBot World 1,001,552 trajectories (2,976.4 h), reporting 32% higher performance than RDT.
Notes: Fig. 5(b) bar values (GO-1 average, RDT-1B, w/o Latent Planner) are not used because their correspondence could not be confirmed in the original text. Affiliation per page 1 of the paper (OpenDriveLab is the GitHub organization name, corrected from the note). Weights HF agibot-world/GO-1 (CC BY-NC-SA 4.0), open-sourced 2025-09-19. No GPU sentence in the paper.
Verification: Corrections applied