RACE_main_gr00t — GR00T N1.6 LIBERO checkpoints

folder	what	action chunk	notes
teacher_h8/	GR00T N1.6-3B fine-tuned on merged LIBERO (4 suites), 30k steps, batch 64	8	starting point for the runs below (original, unmodified)
ft_h16_20k/	teacher_h8 → fine-tuned with chunk 16, 20k steps	16
ft_h24_20k/	teacher_h8 → fine-tuned with chunk 24, 20k steps	24
ft_h32_20k/	teacher_h8 → fine-tuned with chunk 32, 20k steps	32

Fine-tuning recipe (identical across H): stock Isaac-GR00T n1.6-release, embodiment libero_panda with action.delta_indices = range(H), VLM frozen (tune_llm=False, tune_visual=False, tune_top_llm_layers=0), action head trained (tune_projector/tune_diffusion_model/tune_vlln=True), global batch 64 on 2 GPUs, lr 1e-4 cosine with 5% warmup, weight decay 1e-5, DeepSpeed ZeRO-2, bf16, data = IPEC-COMMUNITY/libero_{spatial,object,goal,10}_no_noops_1.0.0_lerobot merged. Checkpoints hold model weights + config + processor (optimizer state removed).

Original LIBERO (50 episodes/task, executing the whole chunk), 20k step:

H	Spatial	Object	Goal	Long	Avg
16	100.0	100.0	98.6	94.9	98.4
24	100.0	100.0	98.4	89.6	97.0
32	98.6	98.8	95.7	90.2	95.8

Serving: gr00t/eval/run_gr00t_server.py --model-path <folder> --embodiment-tag LIBERO_PANDA --use-sim-policy-wrapper; the chunk length is read from the checkpoint's processor config. Note teacher_h8/processor_config.json carries a letter_box_transform key from a fork; delete it to load with stock n1.6.