ERVLA: Embodied Chain-of-Thought for Generalizable Robot Manipulation

Predict a robot action chunk from two camera views + a task instruction, with the LIBERO-Plus-tuned ERVLA checkpoint (ERVLA/liberoplus — Qwen3-VL-4B backbone + a DiT flow-matching action head; 86.9% success on LIBERO-Plus). Give the robot's eye views, tell it what to do, and read the predicted end-effector deltas (Δxyz, Δrotation, gripper). No simulator runs here — the action chunk is the output.

[paper] · code] · model