wkal/omy_insert_manual_qactual_state7_future1

OMY Insert Manual Dual View q_actual State7 Future1

This is the converted dual-view LeRobot dataset for the MuJoCo plug insertion task.

Semantics

• —fps: 20 Hz
• —observation.image: agent-view RGB image, 256 x 256
• —observation.wrist_image: wrist/egocentric RGB image, 256 x 256
• —observation.state[t]: measured joint state q_actual[t], 7D [joint1..joint6, gripper_actual]
• —action[t]: measured next-frame joint state q_actual[t+1], 7D [joint1..joint6, gripper_actual]
• —With chunk_size=15, a training sample at frame t uses: action_chunk_t = [q_actual[t+1], ..., q_actual[t+15]]

The source manual collection stored same-frame measured joint states as action. This dataset shifts those labels by one frame and uses the same measured joint state as the policy state input.

Recommended Training Config

Use smolvla_insert_manual_y_env4_qactual_state7_chunk15.yaml from the companion GitHub repository:

• —chunk_size: 15
• —n_action_steps: 10
• —batch_size: 1
• —observation.state.shape: [7]
• —action.shape: [7]

At inference, the policy predicts 15 actions and executes the first 10 before replanning from a fresh observation.