Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern_MolmoAct2_QuantilesFullRange_bs16_step3000

MolmoAct2 Quantiles Full Range - Task 000004 Peanut Pick and Place

Public LeRobot 0.6.1 inference checkpoint saved at step 3,000, batch size 16. Training used all 35 episodes of Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern at revision 05286a17a145234ed80870702f4d9757f00194c3.

• Cameras: cam_left_head, cam_left_wrist, cam_right_wrist
• State/action dimensions: 22 / 22
• Action chunk and default executed actions: 10 / 10
• State/action normalization: QUANTILES with q01=dataset min and q99=dataset max
• Load the saved policy preprocessor and postprocessor with the model

Open-loop first-action evaluation on 100 uniformly sampled frames from episode 5: MAE 0.00573342, RMSE 0.02800866 in original action units. Episode 5 was part of training, so this is a training-set diagnostic rather than held-out validation or real-robot rollout performance.

Only inference files are published. Optimizer, scheduler, and RNG states remain in the local training checkpoint.