mirror of
https://github.com/huggingface/lerobot.git
synced 2026-07-06 17:41:47 +00:00
johnnynunez
20c0f07858
Isaac-GR00T applies dual state regularization during fine-tuning: raw-state zeroing driven by the processor sidecar's state_dropout_prob (0.2 for the inspected N1.7 checkpoint) plus encoded-feature dropout. Baseline LeRobot kept the processor in deterministic mode, so the raw-state dropout never activated (RCA Tier-2 contributor to the LeRobot-trained SO-101 failures). - GrootN17PackInputsStep: runtime-only 'training' flag + state_dropout_prob; whole-sample state zeroing gated on torch.is_grad_enabled() so eval and no_grad validation paths are unaffected - sidecar loader reads state_dropout_prob from processor_config.json - state_dropout_prob serializes with the step; the training flag intentionally does not (reloaded pipelines default to eval, re-enabled only when processors are rebuilt with dataset_meta) - _set_groot_preprocessor_training toggles any dataclass step exposing a 'training' field on serialized-pipeline reloads Verification: tests/policies/groot/test_groot_state_dropout.py (4 passed) on RTX PRO 6000 / CUDA 13.3.
101 lines
3.2 KiB
Python
101 lines
3.2 KiB
Python
#!/usr/bin/env python
|
|
|
|
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
"""Isaac-GR00T N1.7 raw-state dropout training contract.
|
|
|
|
Isaac-GR00T zeroes the entire proprioceptive state of a sample with probability
|
|
``state_dropout_prob`` (configured in the checkpoint's processor sidecar) during
|
|
training only. Baseline LeRobot kept the processor deterministic, so this
|
|
regularization never activated. These tests pin the train/eval split.
|
|
"""
|
|
|
|
import torch
|
|
|
|
from lerobot.policies.groot.processor_groot import GrootN17PackInputsStep
|
|
from lerobot.types import TransitionKey
|
|
from lerobot.utils.constants import OBS_STATE
|
|
|
|
|
|
def _make_transition():
|
|
return {
|
|
TransitionKey.OBSERVATION: {OBS_STATE: torch.tensor([[1.0, 2.0], [3.0, 4.0]])},
|
|
TransitionKey.COMPLEMENTARY_DATA: {"task": ["Move", "Move"]},
|
|
}
|
|
|
|
|
|
def test_groot_n1_7_training_applies_raw_state_dropout_before_encoder():
|
|
step = GrootN17PackInputsStep(
|
|
max_state_dim=4,
|
|
max_action_dim=4,
|
|
normalize_min_max=False,
|
|
training=True,
|
|
state_dropout_prob=1.0,
|
|
)
|
|
|
|
output = step(_make_transition())
|
|
|
|
expected = torch.zeros(2, 1, 4)
|
|
torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
|
|
|
|
|
|
def test_groot_n1_7_training_state_dropout_is_disabled_under_no_grad():
|
|
step = GrootN17PackInputsStep(
|
|
max_state_dim=4,
|
|
max_action_dim=4,
|
|
normalize_min_max=False,
|
|
training=True,
|
|
state_dropout_prob=1.0,
|
|
)
|
|
|
|
with torch.no_grad():
|
|
output = step(_make_transition())
|
|
|
|
expected = torch.tensor([[[1.0, 2.0, 0.0, 0.0]], [[3.0, 4.0, 0.0, 0.0]]])
|
|
torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
|
|
|
|
|
|
def test_groot_n1_7_eval_mode_state_dropout_is_inactive():
|
|
step = GrootN17PackInputsStep(
|
|
max_state_dim=4,
|
|
max_action_dim=4,
|
|
normalize_min_max=False,
|
|
training=False,
|
|
state_dropout_prob=1.0,
|
|
)
|
|
|
|
output = step(_make_transition())
|
|
|
|
expected = torch.tensor([[[1.0, 2.0, 0.0, 0.0]], [[3.0, 4.0, 0.0, 0.0]]])
|
|
torch.testing.assert_close(output[TransitionKey.OBSERVATION]["state"], expected)
|
|
|
|
|
|
def test_groot_n1_7_pack_step_serializes_dropout_prob_but_not_training_mode():
|
|
step = GrootN17PackInputsStep(
|
|
max_state_dim=4,
|
|
max_action_dim=4,
|
|
normalize_min_max=False,
|
|
training=True,
|
|
state_dropout_prob=0.2,
|
|
)
|
|
|
|
serialized = step.get_config()
|
|
restored = GrootN17PackInputsStep(**serialized)
|
|
|
|
assert "training" not in serialized
|
|
assert serialized["state_dropout_prob"] == 0.2
|
|
assert restored.training is False
|
|
assert restored.state_dropout_prob == 0.2
|