Instructions to use javadcc/evo1-libero-lerobot with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use javadcc/evo1-libero-lerobot with LeRobot:
- Notebooks
- Google Colab
- Kaggle
EVO-1 LIBERO LeRobot Checkpoint
This repository contains the converted LeRobot-format EVO-1 LIBERO checkpoint.
Source checkpoint:
/llm_jzm/EAI_VLA/project/evo1_ckpt_convert/Evo1_LIBERO
Converted checkpoint validation:
- The checkpoint loads with
EVO1Policy.from_pretrained. config.jsonhastype=evo1anduse_flash_attn=true.- Official old EVO-1 checkpoint with FlashAttention enabled:
libero_object, 10 tasks x 1 episode, 10/10 success. - Converted LeRobot checkpoint with FlashAttention enabled:
libero_object, 10 tasks x 1 episode, 10/10 success. - Same LIBERO observation and action seed: pixel, VIT, fused-token, normalized action7, and denormalized action7 max absolute diff are all 0.0 between official old and converted LeRobot paths.
Validation artifacts were produced under /llm_jzm/EAI_VLA/project/evo1_ckpt_convert/.
- Downloads last month
- 37