Modelpublic

marin6670/0915_isaacsim_auto_only_model_200k_steps

sourceHugging Faceupdated 10d agoView on Hugging Face

0likes17downloads

Model Card

0915isaacsimautoonlymodel200ksteps

SmolVLA fine-tuned for 200,000 optimizer steps on the auto-only SO101 Isaac Sim dataset.

Setting	Value
Base model	lerobot/smolvla_base
Dataset	0915_isaacsim_so101_auto_block_basketball_coca_cola_samsung_tv_remote_control_dataset
Training episodes	200; 50 per target
Batch size	8
Seed	1000
Vision encoder	Frozen
Train expert only / state projection	True / True
Learning rate	0.0001; 1,000 warmup steps; cosine decay to 0.0000025
Chunk size / action steps	50 / 50
Camera mapping	front→camera1, top→camera2, wrist→camera3

The final 200,000-step inference checkpoint and preprocessing/postprocessing statistics are included. Optimizer and RNG resume state remain with the original local training run. The checkpoint weights are unchanged; repository/job references use the publication names. The original training configuration is preserved under provenance/.

Evaluation results are pending. No success rate is claimed in this initial publication.