OpenVLA Sim2Real T1 Spatial ACSSP Step 1000
Merged OpenVLA-OFT checkpoint for SafeVLA/SafeLIBERO T1 spatial sim2real adaptation.
- Initial checkpoint:
/mnt/dataset/kai_data/saferl_ckp/safelibero_spatial_task0_acssp_obsact_lam020_hinge05_step39
- Adaptation data:
/mnt/dataset/safevla_t1_spatial_success/tfds_standard_success, dataset safelibero_spatial_rawfull
- Training run id:
sim2real_t1_standard_success_longtrend_from_acssp_task0_lam020_hinge05_step39_lr5e-6_s2000_r8
- Selected checkpoint: step 1000
- Training config: LoRA rank 8, LR 5e-6, batch size 8/GPU, 8 GPUs, image augmentation enabled, discrete-token SFT
This repository includes the merged model shards plus the lora_adapter/ saved by the training script.
