Gemini Robotics On-Device 2

Google DeepMind · 2026-07

VLANew in 2026Not disclosedNot disclosed

A VLA that runs locally on the robot, adapting to new bimanual embodiments in a few hours·with under 200 examples.

Paper

Backbone LLM: Gemma-based (size not disclosed)
Vision encoder: —
Action head: Numeric robot actions (Not disclosed)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): On-device Gemma model + Gemini Robotics 1.5 technology; input text·images·proprioception, output numeric robot actions

Pretraining accelerators: TPU (TPU Pods, JAX·ML Pathways; generation·count not disclosed)
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: Adaptation takes a few hours, usually fewer than 200 examples; hardware not disclosed
Data type: Web vision-language (images·text) + robot sensor·action data
Data scale: Not disclosed
Main datasets: Not disclosed
Tasks & evaluation: Post-training progress on new platforms: SO101 6.7%→53.3% (v1 0.0%→6.7%), Dexmate 24.4%→75.6% (v1 13.3%→33.3%)
Robot: SO101, Dexmate, bimanual robot

Inference device & speed: Runs locally on robot; hardware·latency Not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Onboard

Reinforcement learning: Not disclosed · Training method details not disclosed
Base control: Fixed · No base control (bimanual target; mobile platforms·whole-body control explicitly out of scope)

Key features: A VLA that runs locally on the robot, adapting to new bimanual embodiments in a few hours·with under 200 examples. SO101 success during post-training rose from 6.7% to 53.3% (v1 0.0%→6.7%).
Notes: Corrections applied: 6.7%→53.3% and 24.4%→75.6% are not a v1→v2 comparison but gains from v2's own post-training progress. Card dates conflict: PDF 2026-07-12, web 2026-07-30. The PDF body uses the unnumbered 'Gemini Robotics On-Device'.
Verification: Corrections applied