Preserve native DROID checkpoint behavior for Positronic

.github/workflows/docker-build.yml

@@ -0,0 +1,36 @@
+name: Build and Push Gr00t Images
+
+on:
+  push:
+    branches:
+      - main-positronic
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Free disk space
+        run: |
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/CodeQL"
+          sudo rm -rf "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/node"
+          sudo rm -rf "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/go"
+          sudo rm -rf /usr/lib/jvm
+          sudo docker system prune -af || true
+          df -h
+
+      - name: Log in to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKER_USERNAME }}
+          password: ${{ secrets.DOCKER_PASSWORD }}
+
+      - name: Build and Push
+        working-directory: docker
+        run: |
+          make push

README.md

@@ -255,7 +255,7 @@ See the [Orin setup guide](scripts/deployment/README.md#jetson-orin-setup) for D
 > per-platform Docker and bare-metal setup.
 
 
-For a containerized setup that avoids system-level dependency conflicts, see our [Docker Setup Guide](docker/README.md). The recommended container workflow is to start the image first, then clone or pull the repo inside the running container so your checkout uses the image's prebuilt dependency environment.
+For a containerized setup that avoids system-level dependency conflicts, see our [Docker Setup Guide](docker/README.md). The image includes this fork at `/gr00t` with its prebuilt dependency environment.
 
 ---
 

docker/Dockerfile

@@ -225,4 +225,10 @@ ENV MUJOCO_GL="egl" \
 # bootstrap_wheels.sh before sync (then committed back) if missing. flash-attn
 # comes from official release URLs — no source build needed.
 
+# Install the fork alongside its dependency environment for Positronic subprocesses.
+WORKDIR /gr00t
+COPY . /gr00t
+RUN uv pip install --python /opt/gr00t-venv/bin/python --no-deps -e /gr00t
+ENV PYTHONPATH=/gr00t
+
 CMD ["/bin/bash"]

docker/Makefile

@@ -0,0 +1,79 @@
+.PHONY: all build tag push clean prune help
+
+# Image configuration
+IMAGE_NAME_GROOT := positro/gr00t-base
+
+# Extract version from pyproject.toml (first literal version entry)
+VERSION := $(shell sed -n 's/^version = "\([^"]*\)"/\1/p' ../pyproject.toml | head -n 1)
+
+GIT_SHA := $(shell git rev-parse --short HEAD 2>/dev/null || echo "unknown")
+
+REGISTRY_URL ?= docker.io
+
+TAG_GROOT_LATEST := $(IMAGE_NAME_GROOT):latest
+TAG_GROOT_VERSION := $(IMAGE_NAME_GROOT):v$(VERSION)
+TAG_GROOT_SHA := $(IMAGE_NAME_GROOT):$(GIT_SHA)
+LOCAL_TAG_GROOT := $(IMAGE_NAME_GROOT):local
+
+help:
+	@echo "gr00t Groot Docker Build System"
+	@echo ""
+	@echo "Configuration:"
+	@echo "  Groot Image:   $(IMAGE_NAME_GROOT)"
+	@echo "  Version:       $(VERSION)"
+	@echo "  Git SHA:       $(GIT_SHA)"
+	@echo "  Registry URL:  $(REGISTRY_URL)"
+	@echo ""
+	@echo "Targets:"
+	@echo "  make build     Build the groot image"
+	@echo "  make tag       Tag the groot image"
+	@echo "  make push      Push groot tags to Docker Hub"
+	@echo "  make clean     Remove local groot images"
+	@echo "  make prune     Remove dangling/unused Docker images"
+	@echo "  make help      Show this help message"
+	@echo ""
+
+build:
+	@echo "Building $(IMAGE_NAME_GROOT) image..."
+	@if [ -z "$(VERSION)" ]; then \
+		echo "Error: Could not extract version from pyproject.toml"; \
+		exit 1; \
+	fi
+	docker build --platform linux/amd64 \
+		-f Dockerfile \
+		-t $(LOCAL_TAG_GROOT) ..
+
+tag: build
+	@echo "Tagging groot image with multiple tags..."
+	docker tag $(LOCAL_TAG_GROOT) $(TAG_GROOT_LATEST)
+	docker tag $(LOCAL_TAG_GROOT) $(TAG_GROOT_VERSION)
+	docker tag $(LOCAL_TAG_GROOT) $(TAG_GROOT_SHA)
+	@echo "Tagged with:"
+	@echo "  - $(TAG_GROOT_LATEST)"
+	@echo "  - $(TAG_GROOT_VERSION)"
+	@echo "  - $(TAG_GROOT_SHA)"
+
+push: tag
+	@echo "Pushing groot images to Docker Hub..."
+	docker push $(TAG_GROOT_LATEST)
+	docker push $(TAG_GROOT_VERSION)
+	docker push $(TAG_GROOT_SHA)
+	@echo ""
+	@echo "Successfully pushed groot images to Docker Hub!"
+	@echo "To use on cloud instances, set:"
+	@echo "  export IMAGE_REGISTRY=$(REGISTRY_URL)/"
+
+all: push
+
+clean:
+	@echo "Removing local groot images..."
+	-docker rmi $(LOCAL_TAG_GROOT)
+	-docker rmi $(TAG_GROOT_LATEST)
+	-docker rmi $(TAG_GROOT_VERSION)
+	-docker rmi $(TAG_GROOT_SHA)
+	@echo "Cleanup complete."
+
+prune:
+	@echo "Pruning dangling and unused Docker images..."
+	docker image prune -f
+	@echo "Prune complete."

docker/README.md

@@ -18,11 +18,38 @@ From the repository root:
 bash docker/build.sh
 ```
 
-This builds from `nvidia/cuda:12.8.0-devel-ubuntu24.04` and installs all dependencies into `/opt/gr00t-venv`. The image does not include a working source checkout; for normal use, start the image and then clone or pull the repo you want to run inside the container.
+This builds from `nvidia/cuda:12.8.0-devel-ubuntu24.04` and installs all dependencies into `/opt/gr00t-venv`. The image includes this fork at `/gr00t`, installed into `/opt/gr00t-venv`.
+
+## Positronic base image
+
+```bash
+make -C docker build
+make -C docker push
+```
+
+The Makefile publishes `positro/gr00t-base` with `latest`, version, and commit tags.
+It targets `linux/amd64`, including when built on an Apple Silicon Mac.
+For a native ARM build, use `bash docker/build.sh` on the target host.
+Positronic's `GROOT_BASE_IMAGE` selects an existing base image for its adapter build.
+
+Fine-tuning defaults to the base checkpoint's saved model and modality configuration.
+The Python launcher accepts `--video-keys` to select the fine-tuning camera layout; omitted, it retains the checkpoint's views.
+When using `examples/finetune.sh`, put this option after the script's `--` passthrough delimiter:
+
+```bash
+bash examples/finetune.sh \
+    --base-model-path nvidia/GR00T-N1.7-DROID \
+    --dataset-path /data/droid \
+    --embodiment-tag oxe_droid_relative_eef_relative_joint \
+    --output-dir /data/checkpoints \
+    -- --video-keys exterior_image_1_left exterior_image_2_left wrist_image_left
+```
+
+The policy server accepts `--model-path hf://nvidia/GR00T-N1.7-DROID` and downloads that snapshot.
 
 ## Running the Container
 
-**Recommended workflow: run the image, then clone or update the repo inside it.**
+**Run the included fork from `/gr00t`.**
 
 Start an interactive shell:
 
@@ -35,9 +62,7 @@ docker run -it --rm --gpus all \
 Then, inside the container:
 
 ```bash
-git clone --recurse-submodules https://github.com/NVIDIA/Isaac-GR00T /workspace/Isaac-GR00T
-cd /workspace/Isaac-GR00T
-export PYTHONPATH="$PWD${PYTHONPATH:+:$PYTHONPATH}"
+cd /gr00t
 python -c "import gr00t; print('GR00T ready')"
 ```
 

gr00t/configs/finetune_config.py

@@ -41,10 +41,13 @@ class FinetuneConfig:
 
     modality_config_path: str | None = None
     """
-    Path to a Python file defining the modality configuration for the given embodiment. 
-    If None, use the pre-registered modality config in `gr00t/configs/data/embodiment_configs.py`. 
+    Path to a Python file defining the modality configuration for the given embodiment.
+    If None, retain the modality configuration saved in the base checkpoint.
     """
 
+    video_keys: list[str] | None = None
+    """Explicit camera keys for the fine-tuning dataset; omission retains the checkpoint's cameras."""
+
     # --- Model Tuning Flags ---
     tune_llm: bool = False
     """If True, fine-tune the language model (LLM) backbone during training."""

gr00t/data/types.py

@@ -22,6 +22,10 @@
 from gr00t.data.embodiment_tags import EmbodimentTag
 
 
+VIDEO = "video"
+LANGUAGE = "language"
+
+
 class MessageType(Enum):
     START_OF_EPISODE = "start_of_episode"
     END_OF_EPISODE = "end_of_episode"

gr00t/experiment/launch_finetune.py

@@ -16,15 +16,19 @@
 # Launch finetuning for N1.7 on "single node".
 # This script tries to provide a similar user experience as current OSS.
 
+import copy
 import json
 import os
 from pathlib import Path
 
+from transformers.utils import cached_file
 import tyro
 
 from gr00t.configs.base_config import get_default_config
 from gr00t.configs.finetune_config import FinetuneConfig
-from gr00t.experiment.experiment import run
+from gr00t.configs.model.gr00t_n1d7 import Gr00tN1d7Config
+from gr00t.data.embodiment_tags import EmbodimentTag
+from gr00t.data.types import ModalityConfig
 
 
 # Make sure the user provided modality config is registered.
@@ -41,16 +45,9 @@ def load_modality_config(modality_config_path: str):
         raise FileNotFoundError(f"Modality config path does not exist: {modality_config_path}")
 
 
-if __name__ == "__main__":
-    # Set LOGURU_LEVEL environment variable if not already set (default: INFO)
-    if "LOGURU_LEVEL" not in os.environ:
-        os.environ["LOGURU_LEVEL"] = "INFO"
-    # Use tyro for clean CLI
-    ft_config = tyro.cli(FinetuneConfig, description=__doc__)
-    from gr00t.data.embodiment_tags import EmbodimentTag
-
-    ft_config.embodiment_tag = EmbodimentTag.resolve(ft_config.embodiment_tag)
-    embodiment_tag = ft_config.embodiment_tag.value
+def build_config(ft_config: FinetuneConfig):
+    """Inherit model and modality contracts from the checkpoint before applying training overrides."""
+    embodiment_tag = EmbodimentTag.resolve(ft_config.embodiment_tag).value
 
     # all rank workers should register for the modality config
     if ft_config.modality_config_path is not None:
@@ -73,21 +70,42 @@ def load_modality_config(modality_config_path: str):
         }
     )
     config.load_config_path = None
+    config.model = Gr00tN1d7Config.from_pretrained(ft_config.base_model_path)
+    checkpoint = Path(ft_config.base_model_path)
+    processor_root = checkpoint / "processor" if (checkpoint / "processor").is_dir() else checkpoint
+    processor_file = cached_file(str(processor_root), "processor_config.json")
+    with open(processor_file) as f:
+        processor_kwargs = json.load(f)["processor_kwargs"]
+    config.model.use_relative_action = processor_kwargs["use_relative_action"]
+    if ft_config.modality_config_path is None:
+        modalities = processor_kwargs["modality_configs"][embodiment_tag]
+        config.data.modality_configs = {
+            embodiment_tag: {name: ModalityConfig(**value) for name, value in modalities.items()}
+        }
+    else:
+        config.data.modality_configs = copy.deepcopy(config.data.modality_configs)
+    if ft_config.video_keys is not None:
+        if not ft_config.video_keys or len(set(ft_config.video_keys)) != len(ft_config.video_keys):
+            raise ValueError("video_keys must be nonempty and unique")
+        config.data.modality_configs[embodiment_tag]["video"].modality_keys = ft_config.video_keys
 
     # overwrite with finetune config supplied by the user
     config.model.tune_llm = ft_config.tune_llm
     config.model.tune_visual = ft_config.tune_visual
     config.model.tune_projector = ft_config.tune_projector
     config.model.tune_diffusion_model = ft_config.tune_diffusion_model
     config.model.state_dropout_prob = ft_config.state_dropout_prob
-    config.model.random_rotation_angle = ft_config.random_rotation_angle
-    config.model.color_jitter_params = ft_config.color_jitter_params
+    if ft_config.random_rotation_angle is not None:
+        config.model.random_rotation_angle = ft_config.random_rotation_angle
+    if ft_config.color_jitter_params is not None:
+        config.model.color_jitter_params = ft_config.color_jitter_params
     config.model.use_percentiles = ft_config.use_percentiles
     if (ft_config.shortest_image_edge is None) != (ft_config.crop_fraction is None):
         raise ValueError("shortest_image_edge and crop_fraction must be set together")
     if ft_config.shortest_image_edge is not None:
         config.model.shortest_image_edge = ft_config.shortest_image_edge
         config.model.crop_fraction = ft_config.crop_fraction
+    if config.model.shortest_image_edge is not None and config.model.crop_fraction is not None:
         config.model.image_crop_size = None
         config.model.image_target_size = None
     if ft_config.extra_augmentation_config:
@@ -99,7 +117,6 @@ def load_modality_config(modality_config_path: str):
     config.model.reproject_vision = False
     config.model.model_name = "nvidia/Cosmos-Reason2-2B"
     config.model.backbone_trainable_params_fp32 = True
-    config.model.use_relative_action = True
 
     config.training.experiment_name = ft_config.experiment_name
     config.training.start_from_checkpoint = ft_config.base_model_path
@@ -127,4 +144,11 @@ def load_modality_config(modality_config_path: str):
     config.training.resume_from_checkpoint = ft_config.resume_from_checkpoint
     config.training.skip_weight_loading = ft_config.skip_weight_loading
 
-    run(config)
+    return config
+
+
+if __name__ == "__main__":
+    from gr00t.experiment.experiment import run
+
+    os.environ.setdefault("LOGURU_LEVEL", "INFO")
+    run(build_config(tyro.cli(FinetuneConfig, description=__doc__)))

gr00t/model/gr00t_n1d7/processing_gr00t_n1d7.py

@@ -872,6 +872,8 @@ def from_pretrained(cls, pretrained_model_name_or_path: str | Path, **kwargs):
                 "exclude_state",
                 "state_dropout_prob",
                 "use_mean_std",
+                "use_percentiles",
+                "extra_augmentation_config",
                 "model_name",
                 "model_type",
                 "max_action_horizon",