VLM · local, aws-gpu

Gemma 4

Gemma 4 is the perception/tool VLM path for robot camera streams, sim frames, and multimodal workflow glue.

Runtime Target

Start with VLM frame analysis around MuJoCo or real robot cameras; use cloud when local memory is not enough.

Plan

PRO

Provider

aws-batch

Accelerator

nvidia-l4-or-a10g

Simulator

mujoco-camera-stream

Embodiment

robot camera stream

Placement

local-or-cloud

Status

Working Now

AWS Batch runs a Gemma 4 GPU smoke on L4 with the Hugging Face model id, alongside catalog metadata, paid gating, autoscaling, and session planning.

Ready To Deploy

Scale-to-zero AWS Batch infrastructure, ECR worker image wiring, and cloud-session control-plane responses are deployed as the first production path.

Next To Wire

Production camera-stream serving, longer VLM evaluations, and Nebius lifecycle dispatch still need deployment hardening.