PyTorch conversion of qualiadev/pi05-ringonpole-both-bs64-ah60 via openpi's examples/convert_jax_model_to_pytorch.py (bfloat16), produced by training/recap/convert_sft_checkpoint.py (q-research, QUA-1149).

Config: pi05, action_horizon 60, action_dim 32, discrete_state_input true (state rides in the discretized prompt tokens, as trained).

Norm stats are at both assets/qualiadev/ringonpole-v21/norm_stats.json (openpi serving layout) and qualiadev/ringonpole-v21/norm_stats.json (RLinf get_openpi_config layout).