Upgrade GR00T integration to native N1.7 DROID behavior

README.md

@@ -82,7 +82,7 @@ Positronic supports state-of-the-art foundation models with first-class workflow
 | Model | Capability | Training | Inference | Best For |
 |-------|-----------|----------|-----------|----------|
 | **[OpenPI (π₀.₅)](positronic/vendors/openpi/README.md)** | Most capable, generalist | Capable GPU (~78GB, LoRA) | Capable GPU (~62GB) | Complex multi-task manipulation |
-| **[GR00T](positronic/vendors/gr00t/README.md)** | Generalist robot policy | Capable GPU (~50GB) | Smaller GPU (~7.5GB) | Logistics and industry applications |
+| **[GR00T N1.7 DROID](positronic/vendors/gr00t/README.md)** | DROID robot policy | CUDA GPU | CUDA GPU | Joint control with 2 or 3 camera views |
 | **[LeRobot SmolVLA](positronic/vendors/lerobot/README.md)** | VLM-based, multi-task | Consumer GPU | Consumer GPU | Multi-task manipulation with language |
 | **[LeRobot ACT](positronic/vendors/lerobot_0_3_3/README.md)** | Single-task, efficient | Consumer GPU | Consumer GPU | Specific manipulation tasks |
 

docker/Dockerfile.groot

@@ -1,14 +1,18 @@
-ARG BASE_IMAGE=positro/gr00t-base:latest
+ARG BASE_IMAGE
 FROM ${BASE_IMAGE}
 
-# Set working directory
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /usr/local/bin/
+
 WORKDIR /positronic
 
-# Copy the positronic repository
+RUN apt-get update && apt-get install -y --no-install-recommends libturbojpeg && rm -rf /var/lib/apt/lists/*
+
 COPY . /positronic
 
-# Set Python path to include source root and gr00t
 ENV PYTHONPATH=/positronic:/gr00t
 
-# Default command
+# Keep Positronic's dependencies separate from the checkpoint runtime.
+ENV UV_PROJECT_ENVIRONMENT=/positronic/.venv VIRTUAL_ENV=/positronic/.venv
+RUN --mount=type=cache,target=/root/.cache/uv uv sync --locked --python 3.12 --no-dev
+
 CMD ["bash"]

docker/Makefile

@@ -1,5 +1,7 @@
 .PHONY: all build tag push clean prune help build-training tag-training push-training build-openpi tag-openpi push-openpi build-groot tag-groot push-groot build-dreamzero-base push-dreamzero-base build-dreamzero tag-dreamzero push-dreamzero build-robolab tag-robolab push-robolab nebius-login push-robolab-cr
 
+.DEFAULT_GOAL := help
+
 # Image configuration
 IMAGE_NAME_TRAINING := positro/positronic
 IMAGE_NAME_OPENPI := positro/openpi
@@ -41,6 +43,16 @@ TAG_DREAMZERO_BASE_LATEST := $(IMAGE_NAME_DREAMZERO_BASE):latest
 # OpenPI base image (pulled from registry)
 TAG_OPENPI_BASE := $(IMAGE_NAME_OPENPI_BASE):latest
 
+GROOT_REF := be79d6244dda302ace1ff7a2cd55239aad2ad109
+ifndef GROOT_BASE_IMAGE
+GROOT_BASE_IMAGE := positro/gr00t-base:$(GROOT_REF)
+build-groot: build-groot-base
+endif
+
+.PHONY: build-groot-base
+build-groot-base:
+	docker build --platform linux/amd64 -f docker/Dockerfile -t $(GROOT_BASE_IMAGE) https://github.com/Positronic-Robotics/gr00t.git\#$(GROOT_REF)
+
 help:
 	@echo "Positronic Docker Build System - Makefile"
 	@echo ""
@@ -81,7 +93,7 @@ build-openpi:
 
 build-groot: build-training
 	@echo "Building $(IMAGE_NAME_GROOT)..."
-	docker build --pull --platform linux/amd64 --build-arg BASE_IMAGE=positro/gr00t-base:latest -f Dockerfile.groot -t $(IMAGE_NAME_GROOT):local ..
+	docker build --platform linux/amd64 --build-arg BASE_IMAGE=$(GROOT_BASE_IMAGE) -f Dockerfile.groot -t $(IMAGE_NAME_GROOT):local ..
 
 build-dreamzero-base:
 	@echo "Building $(IMAGE_NAME_DREAMZERO_BASE)..."

docker/README.md

@@ -40,7 +40,23 @@ Now you are ready to build our Docker.
 docker/build.sh
 ```
 
-## GR00T containers: uv mount caveat
+## GR00T N1.7 containers
 
-If you customize `docker-compose.yml` volumes, **do not bind-mount** your host `~/.local/share/uv` into `/root/.local/share/uv` for `positro/gr00t` images.
-GR00T's `/.venv/bin/python` can be a symlink into the image's own uv-managed CPython under `/root/.local/share/uv/python/...`, and the bind mount can hide that target and cause `/.venv/bin/python` to fail with `ENOENT`.
+`make build-groot` builds the GR00T fork revision pinned in `Makefile`. The fork image contains
+CUDA 12.8 and the upstream Python 3.12 environment at `/opt/gr00t-venv`.
+Positronic installs its own locked environment at `/positronic/.venv`.
+Both training and serving launch GR00T in its separate environment.
+
+Build both images from Positronic:
+
+```bash
+make -C docker build-groot
+IMAGE_TAG=local docker compose -f docker/docker-compose.yml run --rm --service-ports groot-server droid
+```
+
+Pass `GROOT_BASE_IMAGE=<image:tag>` to use an existing base and skip its build.
+For local fork development, run `make -C docker build` in the fork,
+then `make -C docker build-groot GROOT_BASE_IMAGE=positro/gr00t-base:local` in Positronic.
+
+Do not mount host uv interpreter directories over the image's interpreter directories.
+See [GR00T](../positronic/vendors/gr00t/README.md) for conversion, fine-tuning and inference.

docker/docker-compose.phail.desktop.yml

@@ -1,13 +1,11 @@
-# GR00T on desktop (RTX 3060 12GB — GR00T uses ~4.6GB)
+# GR00T N1.6 on desktop (RTX 3060 12GB — GR00T uses ~4.6GB)
 #
+# Set GR00T_N16_IMAGE to a pinned N1.6 image.
 # docker --context desktop compose -f docker-compose.phail.desktop.yml up
 services:
   phail-groot-server:
     extends:
-      file: docker-compose.yml
-      service: groot-server
-    container_name: phail-groot-server
-    pull_policy: always
-    command: ["phail"]
+      file: docker-compose.phail.yml
+      service: phail-groot-server
     ports: !override
       - "8000:8000"

docker/docker-compose.phail.yml

@@ -13,6 +13,9 @@ services:
     extends:
       file: docker-compose.yml
       service: groot-server
+    # These experiment checkpoints use the N1.6 action schema.
+    image: ${GR00T_N16_IMAGE:?Set GR00T_N16_IMAGE to a pinned GR00T N1.6 image}
+    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.server"]
     container_name: phail-groot-server
     pull_policy: always
     command: ["phail"]

docker/docker-compose.spoons-ablation.yml

@@ -1,14 +1,17 @@
 # GR00T spoons data ablation servers (~4.6GB each, 2 fit on desktop, 1 on notebook)
 #
 # Desktop (2 servers):
-#   IMAGE_TAG=latest CACHE_ROOT=/home/<user> docker --context desktop compose -f docker-compose.spoons-ablation.yml up spoons-100 spoons-50
+#   GR00T_N16_IMAGE=positro/gr00t:<N1.6-tag> CACHE_ROOT=/home/<user> docker --context desktop compose -f docker-compose.spoons-ablation.yml up spoons-100 spoons-50
 # Notebook (1 server):
-#   IMAGE_TAG=latest docker --context notebook compose -f docker-compose.spoons-ablation.yml up spoons-25
+#   GR00T_N16_IMAGE=positro/gr00t:<N1.6-tag> docker --context notebook compose -f docker-compose.spoons-ablation.yml up spoons-25
 services:
   spoons-100:
     extends:
       file: docker-compose.yml
       service: groot-server
+    # These experiment checkpoints use the N1.6 action schema.
+    image: ${GR00T_N16_IMAGE:?Set GR00T_N16_IMAGE to a pinned GR00T N1.6 image}
+    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.server"]
     container_name: spoons-100
     pull_policy: always
     command:
@@ -21,6 +24,9 @@ services:
     extends:
       file: docker-compose.yml
       service: groot-server
+    # These experiment checkpoints use the N1.6 action schema.
+    image: ${GR00T_N16_IMAGE:?Set GR00T_N16_IMAGE to a pinned GR00T N1.6 image}
+    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.server"]
     container_name: spoons-50
     pull_policy: always
     command:
@@ -33,6 +39,9 @@ services:
     extends:
       file: docker-compose.yml
       service: groot-server
+    # These experiment checkpoints use the N1.6 action schema.
+    image: ${GR00T_N16_IMAGE:?Set GR00T_N16_IMAGE to a pinned GR00T N1.6 image}
+    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.server"]
     container_name: spoons-25
     pull_policy: always
     command:

docker/docker-compose.yml

@@ -181,10 +181,7 @@ services:
     volumes:
       - ${CACHE_ROOT:-${HOME}}/.cache:/root/.cache
       - ${CACHE_ROOT:-${HOME}}/.aws:/root/.aws:ro
-      # NOTE: Do NOT mount `/root/.local/share/uv` for gr00t images.
-      # gr00t's `/.venv/bin/python` is a symlink into the image's own uv-managed CPython under
-      # `/root/.local/share/uv/python/...`. Bind-mounting the host uv dir can hide that target and
-      # make `/.venv/bin/python` fail with ENOENT.
+      # Keep image-owned uv interpreters visible; do not mount /root/.local/share/uv here.
 
     # Set HOME inside container to match the mount expectations
     environment:
@@ -201,12 +198,12 @@ services:
     container_name: groot-train
     shm_size: 8g
     ipc: host  # Enable shared memory access for training
-    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.train"]
+    entrypoint: ["uv", "run", "--no-sync", "--python", "3.12", "python", "-m", "positronic.vendors.gr00t.train"]
 
   groot-server: &groot-server-common
     <<: *groot-common
     container_name: groot-server
-    entrypoint: ["uv", "run", "--python", "3.13", "python", "-m", "positronic.vendors.gr00t.server"]
+    entrypoint: ["uv", "run", "--no-sync", "--python", "3.12", "python", "-m", "positronic.vendors.gr00t.server"]
     ports:
       - "8000:8000"
 

docs/codecs.md

@@ -57,7 +57,7 @@ It is declared in two places, for the two things it does:
 - **Training** — `compose(ee_frame=DROID_EE_FRAME)` re-expresses the dataset in that frame, which is what makes the resulting checkpoint speak it. It defaults to unset, which trains in `default`.
 - **Serving** — the OpenPI pipeline's `ee_frame=` puts the codec left of the `remote` marker, so the rig converts and the server stays frame-agnostic. It has no default: every deployment states its frame — `None` for a checkpoint trained in `default`, or one that speaks joints, which are unambiguous. Nothing checks a stated frame against how the checkpoint was trained, so it is set beside the checkpoint path it belongs to.
 
-Both take the transform itself — `models.DROID_EE_FRAME` is the one we ship — so a checkpoint declares its own frame and no robot model is consulted to serve it. The other vendor servers take no `ee_frame`: every checkpoint they serve was trained in the rig's `default`, so none has a transform to declare.
+Both take the transform itself — `models.DROID_EE_FRAME` is the one we ship — so a checkpoint declares its own frame and no robot model is consulted to serve it. GR00T's DROID codec uses `DROID_EE_FRAME` for both dataset conversion and serving. Its pipeline exposes this through `codec.ee_frame`.
 
 A `CartesianDelta` is the one command this cannot convert on its own: a delta has no anchor pose, so it carries `frame` and the driver composes it where the measured pose lives.
 
@@ -70,7 +70,9 @@ Two wrappers in [`positronic/cfg/codecs.py`](../positronic/cfg/codecs.py) apply
 | Wrapper | Expands to | Used by |
 |---------|-----------|---------|
 | `droid_execution(action)` | `SetControlMode(DROID_IMPEDANCE) \| action` ([the DROID gains](../positronic/cfg/hardware/roboarm/__init__.py)) | the `droid` pipelines of OpenPI, DreamZero and MolmoAct2, and OpenPI's `droid_jointpos` |
-| `phail_v1_execution(action)` | `SetControlMode(PositionControl()) \| action` | the `phail_v1` pipelines of LeRobot, GR00T, OpenPI and DreamZero |
+| `phail_v1_execution(action)` | `SetControlMode(PositionControl()) \| action` | the `phail_v1` pipelines of LeRobot, OpenPI and DreamZero |
+
+GR00T's DROID codec sets `DROID_IMPEDANCE` directly on its joint-position commands.
 
 ## Writing custom codecs
 
@@ -118,31 +120,15 @@ cd docker && docker compose run --rm lerobot-convert convert \
 
 See [`positronic/vendors/gr00t/codecs.py`](../positronic/vendors/gr00t/codecs.py).
 
-| Codec | Observation | Action | Modality Configs |
-|-------|-------------|--------|------------------|
-| `ee_quat` | EE pose (quat) + grip + images (224x224) | Absolute EE position (quat) + grip | `ee`, `ee_rel` |
-| `ee_rot6d` | EE pose (rot6d) + grip + images | Absolute EE position (rot6d) + grip | `ee_rot6d`, `ee_rot6d_rel` |
-| `ee_quat_joints` | EE pose + joints + grip + images | Absolute EE position + grip | `ee_q`, `ee_q_rel` |
-| `ee_rot6d_joints` | EE pose (rot6d) + joints + grip + images | Absolute EE position (rot6d) + grip | `ee_rot6d_q`, `ee_rot6d_q_rel` |
-| `ee_quat_traj` | EE pose (quat) + grip + images | Absolute EE trajectory (quat) + grip (binarized) | `ee`, `ee_rel` |
-| `ee_rot6d_traj` | EE pose (rot6d) + grip + images | Absolute EE trajectory (rot6d) + grip (binarized) | `ee_rot6d`, `ee_rot6d_rel` |
-| `ee_quat_joints_traj` | EE pose + joints + grip + images | Absolute EE trajectory + grip (binarized) | `ee_q`, `ee_q_rel` |
-| `ee_rot6d_joints_traj` | EE pose (rot6d) + joints + grip + images | Absolute EE trajectory (rot6d) + grip (binarized) | `ee_rot6d_q`, `ee_rot6d_q_rel` |
-| `joints_traj` | Joints + grip + images (no EE pose) | Absolute joint trajectory + grip (binarized) | — |
-
-The codec must match the modality config during training.
-
-```bash
-# Convert with codec
-cd docker && docker compose run --rm lerobot-0_3_3-convert convert \
-  --dataset.codec=@positronic.vendors.gr00t.codecs.ee_rot6d_joints \
-  --output_dir=~/datasets/groot/my_task
+| Codec | Cameras | State and training actions | Inference actions |
+|-------|---------|----------------------------|-------------------|
+| `droid` | Exterior + wrist | Absolute EEF pose (XYZ + row-based rot6d), gripper, 7 joints | Absolute joint targets + binary gripper |
+| `droid_three_cameras` | Two exteriors + wrist | Same as `droid` | Same as `droid` |
 
-# Train with matching modality
-cd docker && docker compose run --rm groot-train \
-  --modality_config=ee_rot6d_q \
-  --input_path=~/datasets/groot/my_task
-```
+Images use the upstream DROID client's 320×180 padded resize, then the checkpoint's native
+preprocessing. GR00T owns pose-relative and joint-relative conversion. Training labels are
+recorded state trajectories. Use the same camera layout for conversion and inference;
+the published DROID checkpoint uses two cameras. See the [Docker workflow](../positronic/vendors/gr00t/README.md).
 
 ### OpenPI
 

docs/inference.md

@@ -6,7 +6,7 @@ Deploy trained policies for evaluation and production use. Positronic supports l
 
 Positronic's unified WebSocket protocol connects any hardware to any model (LeRobot, GR00T, OpenPI). The key benefit is running heavy models on powerful GPU hardware (OpenPI needs ~62GB, GR00T ~8GB) separate from the robot/simulator machine.
 
-Each server carries a full **policy pipeline** — one chain naming the rig-side stack, the `remote` split marker, the server-side codec, and the model source that loads checkpoints (see `positronic.policy.spec`). The server runs the half right of the marker and declares the half left of it in its handshake; the client builds the declared stack automatically. Vendors ship their pipelines by name, and every name is a server subcommand — `groot-server ee_rot6d_joints` launches that one. The available names are listed in each vendor's README.
+Each server carries a full **policy pipeline** — one chain naming the rig-side stack, the `remote` split marker, the server-side codec, and the model source that loads checkpoints (see `positronic.policy.spec`). The server runs the half right of the marker and declares the half left of it in its handshake; the client builds the declared stack automatically. Vendors ship their pipelines by name, and every name is a server subcommand — `groot-server droid` launches that one. The available names are listed in each vendor's README.
 
 **Start inference server:**
 ```bash
@@ -20,8 +20,8 @@ cd docker && docker compose run --rm --service-ports lerobot-0_3_3-server ee \
   --pipeline.source.checkpoints_dir=~/checkpoints/lerobot/experiment_v1/
 
 # GR00T
-cd docker && docker compose run --rm --service-ports groot-server ee_rot6d_joints \
-  --pipeline.source.checkpoints_dir=~/checkpoints/groot/experiment_v1/
+cd docker && docker compose run --rm --service-ports -v "$PWD/groot-data:/data" groot-server droid \
+  --pipeline.source.checkpoints_dir=/data/checkpoints/experiment_v1/
 
 # OpenPI (--pipeline.ee_frame states the EE frame the checkpoint speaks; None means the rig's `default`)
 cd docker && docker compose run --rm --service-ports openpi-server ee \

docs/model-selection.md

@@ -11,9 +11,9 @@ Positronic supports three foundation models with different capabilities and reso
 | Aspect | OpenPI (π₀.₅) | GR00T | SmolVLA | LeRobot ACT |
 |--------|---------------|-------|---------|-------------|
 | **Capability** | Most capable, generalist | Generalist | Vision-language-action | Single-task specialist |
-| **Training Hardware** | capable cloud GPU (~78GB, LoRA) | capable cloud GPU (~50GB) | Consumer GPU (RTX 3090, 4090) | Consumer GPU (RTX 3090, 4090) |
+| **Training Hardware** | capable cloud GPU (~78GB, LoRA) | CUDA GPU (measure for batch/layout) | Consumer GPU (RTX 3090, 4090) | Consumer GPU (RTX 3090, 4090) |
 | **Training Time** | Multiple days | 0.5-2 days | Several hours | Several hours |
-| **Inference Hardware** | GPU (~62GB, likely cloud) | GPU (~7.5GB, can run on robot) | Consumer GPU (4GB+) | Consumer GPU (4GB+) |
+| **Inference Hardware** | GPU (~62GB, likely cloud) | CUDA GPU (measure for camera layout) | Consumer GPU (4GB+) | Consumer GPU (4GB+) |
 | **Inference Speed** | Moderate | Moderate | Moderate | Fast |
 | **Best For** | Complex multi-task manipulation, generalization | General robotics tasks | Language-conditioned manipulation | Specific manipulation tasks, fast iteration |
 | **When to Use** | Need generalization, multi-task scenarios, leveraging foundation models | Prefer NVIDIA stack | Language instructions, VLM backbone | Single task, resource constraints, rapid experimentation |
@@ -36,20 +36,14 @@ Positronic supports three foundation models with different capabilities and reso
 
 → [OpenPI Documentation](../positronic/vendors/openpi/README.md)
 
-### GR00T — NVIDIA's Generalist Robot Policy
+### GR00T N1.7 DROID
 
-**What it is:** NVIDIA's foundation model for generalist robot control.
-
-**Strengths:**
-- Generalist capabilities
-- Can run on smaller GPU (~7.5GB inference, can run closer to robot)
-- Requires ~50GB for training (less than OpenPI)
-- Training takes 1-2 days (faster than OpenPI)
-
-**Limitations:**
-- Requires capable GPU for training
-- Slower than single-task models
+NVIDIA's robot policy, using the published DROID checkpoint. Supports wrist plus one exterior
+camera by default, and wrist plus two exteriors through fine-tuning. It predicts joint-position
+actions and preserves the checkpoint's image and relative-action processing.
 
+Training and inference require a CUDA GPU. Measure memory requirements with the intended batch
+size and camera layout.
 
 → [GR00T Documentation](../positronic/vendors/gr00t/README.md)