Work

Vision

Gemma-Le: VLA Policy

I implemented a compact vision-language-action policy in a LeRobot fork, using SigLIP for vision, Gemma 3 with LoRA for language, and ScaleDP for action generation. The checkpoints were trained on the robot_sim.PickNPlace simulation dataset in LeRobot format.

Gemma-Le architecture with an action trajectory being denoised over 50 diffusion steps.
Fig. 1The signal path from multimodal inputs to the ScaleDP action head. The 50-step path is drawn, not a logged rollout. Architecture diagram. The moving trajectory is drawn, not a capture.
Stack
Robotics, VLA, Gemma 3, Diffusion Policy
Links
Hugging Face

Also