Dongkkka/rldx-1-cyclo-intelligence-test

RLDX-1 Cyclo Intelligence Test Fine-tune

Inference-only export of an RLDX-1 fine-tune on Dongkkka/cyclo_intelligence_test_dataset_lerobot_v2.1.

Training Summary

• Base model: RLWRLD/RLDX-1-PT
• Dataset: Dongkkka/cyclo_intelligence_test_dataset_lerobot_v2.1
• Dataset size used: 31 episodes / 5,565 frames
• Camera input: observation.images.cam_head_left only
• Task text: Pick up the green ball and place it into the white basket.
• Training steps: 1,000
• Global batch size: 208
• Final logged step loss: 0.0073
• Mean train loss: 0.022565414

Included Files

This repository intentionally contains only the files needed for inference:

• config.json
• model-00001-of-00004.safetensors
• model-00002-of-00004.safetensors
• model-00003-of-00004.safetensors
• model-00004-of-00004.safetensors
• model.safetensors.index.json
• processor/embodiment_id.json
• processor/processor_config.json
• processor/statistics.json

Checkpoint resume files, optimizer state, scheduler state, RNG state, trainer state, training args, wandb config, and experiment config files are excluded.

Notes

The dataset has 31 available episodes, so the run used all available episodes for a 1,000-step test fine-tune. The model was smoke-tested by loading it with RLDXPolicy and generating action outputs for the left head camera setup.