Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern_MultiTask-DiT-Diffusion_bs96_step30000

MultiTask DiT Diffusion - Task 000004 Peanut Pick & Place

LeRobot 0.6.1 checkpoint at step 30,000, trained on all 35 episodes of Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern (revision 05286a17a145234ed80870702f4d9757f00194c3). There was no held-out validation set or intermediate evaluation.

• Cameras: cam_left_head, cam_left_wrist, cam_right_wrist
• State and action: 22 dimensions each
• Batch size: 96, 30,000 updates
• DiT: 4 layers, hidden size 512, 8 attention heads
• Objective: diffusion with DDPM; 100 training noise steps, 10 inference denoising steps
• Action horizon: 32; default executed actions per call: 24
• Image preprocessing: resize to 320x240, center crop to 224x224 at inference
• State/action normalization: MIN_MAX; visual normalization: MEAN_STD

Load with LeRobot's MultiTaskDiTPolicy.from_pretrained and use the saved preprocessor and postprocessor. The original-unit 22D action is recovered by the postprocessor. Inputs must use the same feature names and three cameras.

This repository contains only the inference checkpoint; optimizer state is not included. Open-loop plots on training episodes are not validation results.