Dongkkka/wallx_dashboard_0904_full_chunk32_10k_8bs

wall_x: final step 10000, batch 8

Final checkpoint, not selected by validation. No intermediate evaluations. Native LeRobot policy training in the Cyclo Intelligence Docker container. Batch size 8; 24 training episodes, 4 held out; 3 cameras; 22D actions; chunk 32. Backbone training enabled using the native policy implementation. WALL-X uses local LeRobot compatibility fixes for 22D action/state projections, strict base-weight loading and vision gradient checkpointing; inference needs the same dimension adapter. Dataset: Dongkkka/cyclodashboard0904testv30. Load with the standard policy loader and the saved pre/post processors. No claim of measured simulator success or evaluation quality.