mimiminsoo/pi0_marker100_real_ft

Continual LoRA fine-tune of mimiminsoo/pi0-lora-marker100-h30 (itself a LoRA fine-tune of pi0_base, trained on 97 sim episodes) further adapted to 30 real-robot SO101 marker-in-cup episodes (mimiminsoo/marker_real_v21, 18,433 frames).

Norm stats are reused from the sim (marker_100) training run rather than recomputed on the real data, so the input/output scale stays consistent with the base LoRA weights.

• config: pi0_marker100_real_ft
• base checkpoint: mimiminsoo/pi0-lora-marker100-h30
• checkpoint step: 2999 / 3000
• lr schedule: cosine decay, peak_lr=1e-5, warmup=100 steps, decay_lr=1e-6
• checkpoints saved every 500 steps under the run for comparison on the real robot

Serving

uv run scripts/serve_policy.py --port=8000 \
    policy:checkpoint --policy.config=pi0_marker100_real_ft --policy.dir=<local path to this repo>