Dongkkka/smolvla_dashboard_0904_20k_64bs

Batch size: 64

Steps: 20000

Full fine-tuned policy checkpoint, including saved pre/post processors and normalization statistics. No adapter merge is required.

Load with SmolVLAPolicy.from_pretrained("Dongkkka/smolvla_dashboard_0904_20k_64bs") and make_pre_post_processors(policy.config, pretrained_path="Dongkkka/smolvla_dashboard_0904_20k_64bs") in a compatible LeRobot environment with SmolVLA dependencies.

For Cyclo Brain, download the complete repository and select SmolVLA with that local model directory. Match the training task instruction, four camera inputs, and 22-dimensional state/action joint mapping. Action chunk size is 50. Initial loading also needs the configured SmolVLM tokenizer/configuration, or a populated local cache.

Validated with strict checkpoint loading and one real dataset observation through saved preprocessing, action inference, and postprocessing. Physical robot execution and task success have not been evaluated by this check.