rosetta_ws

A ready-to-use devcontainer workspace for Rosetta, the ROS2 to LeRobot bridge.

Why This Workspace?

Getting ROS2 and LeRobot installed together is not trivial. This workspace aims to provide an a working example so you can clone, open in VS Code, and get started.

Based on Allison Thackston's excellent ROS2 VS Code devcontainer template (blog post).

Quick Start

git clone https://github.com/iblnkn/rosetta_ws.git
cd rosetta_ws
code .  # VS Code: "Reopen in Container"

The devcontainer:

• Installs ROS2, LeRobot, and all dependencies
• Builds and sources the workspace
• Mounts credentials from host (HuggingFace, W&B, SSH)
• Configures GPU access (NVIDIA runtime)

Workspace Structure

rosetta_ws/
├── src/action/              # ROS2 packages (from repos/src.repos)
│   ├── rosetta/             # Core package - see rosetta/README.md
│   ├── rosetta_interfaces/  # ROS2 action/service definitions
│   ├── lerobot_robot_rosetta/
│   ├── lerobot_teleoperator_rosetta/
│   └── rosetta_rl/
├── libs/                    # Python libraries (from repos/libs.repos)
│   └── lerobot/             # LeRobot (editable install)
├── models/                  # Trained policies
│   ├── act/
│   └── pi05/
├── datasets/
│   ├── bags/                # Raw rosbag recordings
│   └── lerobot/             # Converted LeRobot datasets
├── repos/                   # VCS import definitions
│   ├── src.repos            # ROS2 package sources
│   └── libs.repos           # Python library sources
├── scripts/                 # Build and workflow scripts
├── docker/                  # Dockerfiles (x86, Jetson)
└── .devcontainer/           # VS Code devcontainer configs
    ├── x86/
    └── jetson/

Directories

src/action/

ROS2 packages. Currently embedded in the workspace for development. In production, these will be pulled from separate repositories via repos/src.repos:

# repos/src.repos (not yet active)
repositories:
  action/rosetta:
    type: git
    url: https://github.com/iblnkn/rosetta.git
    version: main

For documentation on the packages themselves, see rosetta/README.md.

libs/

Python libraries installed in editable mode. LeRobot is cloned here so you can modify it if needed.

# repos/libs.repos
repositories:
  lerobot:
    type: git
    url: https://github.com/huggingface/lerobot.git
    version: main

The COLCON_IGNORE file prevents colcon from treating this as a ROS2 package.

models/

Store trained policies here. Organized by policy type:

models/
├── act/
│   └── act_pen_in_cup/
│       └── 050000/
│           └── pretrained_model/
└── pi05/
    └── my_pi05_policy/

Reference in launch files:

ros2 launch rosetta rosetta_client_launch.py \
    pretrained_name_or_path:=/workspaces/rosetta_ws/models/act/my_policy/pretrained_model

Or upload to HuggingFace Hub and reference by repo ID.

datasets/

Two subdirectories for the recording → training pipeline:

Directory	Contents	Created By
datasets/bags/	Raw rosbag2 recordings (MCAP)	Episode Recorder
datasets/lerobot/	Converted LeRobot datasets	port_bags.py

Workflow:

1. Record episodes → datasets/bags/my_robot_YYYYMMDD_HHMMSS/
2. Convert → datasets/lerobot/my_dataset/
3. Train → models/act/my_policy/

repos/

VCS import files for vcs import. To pull external sources:

vcs import src < repos/src.repos
vcs import libs < repos/libs.repos

Currently commented out since packages are embedded for development.

scripts/

Workflow automation scripts:

Script	Description
build.sh	Build workspace with colcon
setup.sh	First-time setup (rosdep, pip installs)
convert_bags_parallel.sh	Parallel bag-to-dataset conversion
train_policy.sh	Training with policy selection and multi-GPU
test.sh	Run tests

Docker

Two Dockerfiles for different platforms:

File	Platform	Base Image
docker/Dockerfile.x86	x86_64 with NVIDIA GPU	nvidia/cuda:12.x + ROS2
docker/Dockerfile.jetson	NVIDIA Jetson (ARM64)	dustynv/ros:jazzy-...

Devcontainer (Recommended)

Open in VS Code and select "Reopen in Container". The devcontainer:

• Uses the appropriate Dockerfile for your platform
• Mounts the workspace at /workspaces/rosetta_ws
• Shares credentials from host (~/.cache/huggingface, ~/.config/wandb, ~/.ssh)
• Enables GPU passthrough

VS Code Tasks

The workspace includes pre-configured tasks (.vscode/tasks.json):

Task	Description
build	Build with colcon
convert bags to lerobot	Convert recordings to dataset
convert bags to lerobot (parallel)	Parallel conversion (faster)
train policy	Interactive policy training
resume training	Resume from checkpoint
upload trained policy	Push to HuggingFace Hub

Access via Ctrl+Shift+P → "Tasks: Run Task".

Credentials

The devcontainer mounts credentials from your host machine:

Host Path	Container Path	Purpose
~/.cache/huggingface	/home/ros/.cache/huggingface	HuggingFace Hub token
~/.config/wandb	/home/ros/.config/wandb	Weights & Biases
~/.netrc	/home/ros/.netrc	Git credentials
~/.ssh	/home/ros/.ssh	SSH keys (read-only)

Login on your host before opening the container:

huggingface-cli login
wandb login

Building

# First time setup
./scripts/setup.sh

# Build packages
./scripts/build.sh
# or
colcon build --symlink-install

# Source
source install/setup.bash

License

Apache-2.0