Gemini Robotics

Google DeepMind · 2025-03

VLAIndustry modelsNot disclosedNot disclosed

Splits into a cloud backbone and an onboard decoder, taking about 250ms from observation to action chunk and controlling at an effective 50Hz.

Paper

Backbone LLM: Based on Gemini 2.0 (size not disclosed)
Vision encoder: Not disclosed
Action head: Onboard local action decoder (Not disclosed)
Chunk: Chunks (length not disclosed)
Control rate: Effective 50 Hz
Architecture (original summary): Gemini 2.0-based VLA backbone (distilled from Gemini Robotics-ER, cloud) + onboard local action decoder; outputs action chunks (chunk length not disclosed)

Pretraining accelerators: TPU v4, v5p, v6e (chip count not disclosed)
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Accelerator·time not disclosed (specialization: 2,000~5,000 episodes per task; fast adaptation: up to 100 demos per task)
Data type: Real-robot teleoperation + web vision-language (documents·code·images·audio·video, ER·VQA)
Data scale: 'thousands of hours' collected over 12 months by an ALOHA 2 fleet, 'thousands of diverse tasks'; exact hours and episode count not disclosed
Main datasets: In-house ALOHA 2 teleoperation data (private) + non-action web data
Tasks & evaluation: 85-task generalization benchmark (in-distribution 20%, visual 28%, instruction 28%, action 24%); specialized long-horizon tasks average 79% (lunch-box 100%); over 70% on 7 of 8 tasks with 100 or fewer demos; bi-arm Franka in-distribution average 63%
Robot: ALOHA 2 (main platform); adapted to bi-arm Franka and Apptronik Apollo humanoid

Inference device & speed: Cloud backbone query-response under 160ms + onboard decoder; observation→action chunk end-to-end about 250ms; effective control 50Hz; onboard computer specs not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Cloud + onboard

Reinforcement learning: Not disclosed · Training method details not disclosed
Base control: Fixed · No base·leg control (ALOHA 2 bimanual; Apollo humanoid used for manipulation tasks only, no mention of locomotion)

Key features: Splits into a cloud backbone and an onboard decoder, taking about 250ms from observation to action chunk and controlling at an effective 50Hz. With 2,000~5,000 episodes per task it can specialize in long-horizon precision tasks such as origami and packing lunch boxes.
Notes: Model card (App. A) Hardware row 'TPU v4, v5p and v6e.', Software row 'JAX, ML Pathways'; the model card covers both Gemini Robotics and Gemini Robotics-ER (corrects the notes' Not disclosed†). Trusted testers per blog: Agile Robots, Agility Robotics, Boston Dynamics, Enchanted Tools. arXiv v1 2025-03-25.
Verification: Corrections applied