DeepThinkVLA — reason, then act
A 3B PaliGemma-based Vision-Language-Action model trained with SFT + RL so that its chain-of-thought actually helps the action it emits. Give it a tabletop scene (agent view + wrist camera) and a task; it writes out its reasoning, then predicts the next 10-step, 7-DoF action chunk in one non-autoregressive pass.
Reading the output. Actions are LIBERO OSC_POSE commands: three normalized
end-effector position deltas, three axis-angle rotation deltas, and a binary gripper
command (+1 closing, -1 opening). At full scale one step is roughly 5 cm / 0.5 rad.
About the images. DeepThinkVLA is trained on LIBERO renders that are rotated 180° by the standard OpenVLA data pipeline, so the example frames look mirrored — that is exactly what the policy expects. Feeding it ordinary photographs is out of distribution.
Example frames come from the authors' yinchenghust/libero_cot
dataset (Apache-2.0). Model code vendored from OpenBMB/DeepThinkVLA (MIT).