admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-23 09:46:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

groot: auto-enable LIBERO gripper action transform for libero_sim

Browse Source 
GR00T N1.7 emits gripper in [0,1] but LIBERO expects [-1,1]. The decode
transform existed but was never auto-enabled for embodiment_tag=libero_sim,
so the policy scored 0% on LIBERO eval. Auto-set it in __post_init__ (still
overridable). LIBERO Spatial eval: 0% -> 98%.
This commit is contained in:
groot-validation
2026-06-05 00:56:11 +00:00
committed by Andy Wrenn
parent 1b24d7bc86
commit 3eb7e0b65b
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/policies/groot/configuration_groot.py
+9
View File
@@ -410,6 +410,15 @@ class GrootConfig(PreTrainedConfig):
if self.base_model_path is None: if self.base_model_path is None:
self.base_model_path = GROOT_N1_7_BASE_MODEL self.base_model_path = GROOT_N1_7_BASE_MODEL
# The N1.7 LIBERO checkpoints emit a [0, 1] gripper action, but the LIBERO
# simulator expects the OpenVLA/[-1, 1] sign convention. NVIDIA's rollout
# wrapper applies this conversion; mirror it here so eval on the
# 'libero_sim' embodiment grasps correctly instead of scoring 0% success.
# This matches the embodiment-specific handling already done for the
# action execution horizon (see infer_groot_n1_7_action_execution_horizon).
if self.action_decode_transform is None and self.embodiment_tag == "libero_sim":
self.action_decode_transform = GROOT_ACTION_DECODE_TRANSFORM_LIBERO
if self.max_state_dim == 64: if self.max_state_dim == 64:
self.max_state_dim = 132 self.max_state_dim = 132
if self.max_action_dim == 32: if self.max_action_dim == 32: