JA | EN

SOBITS VLA Tools

SOBITS VLA Tools is a monorepo providing the full pipeline for controlling SOBITS-developed robots with Vision-Language-Action (VLA) models — from data collection through training to real-time deployment, all integrated via ROS 2. See CONTRIBUTING.md for the conventions new code must follow.

(back to top)

Packages

Package	Purpose	README
sobits_vla_common	Shared hub: robot descriptors, param-schema loader, lerobot compat seam, gamepad client, world reset	README
sobits_vla_rosbag_collection	Gamepad-triggered rosbag recording (C++)	README
sobits_vla_rosbag_conversion	Rosbags → LeRobot dataset	README
sobits_vla_training	Trains/fine-tunes VLA policies via lerobot	README
sobits_vla_deploy	Real-time VLA inference + offline eval	README
sobits_vla_visualization	Reserved for future debug/viz nodes (empty skeleton)	README

All pipeline stages read robot morphology from one robot descriptor (sobits_vla_common/robots/<robot_id>.robot.yaml) — the single source of truth for joint groups, command topics, sensors, and mobile base. Scaffold a new one with:

ros2 run sobits_vla_common new_robot \
  --robot_id sobit_mini --dof 7 --cameras head,hand_left --mobile_base diff \
  --gen_collection_config
ros2 run sobits_vla_common new_robot --robot_id sobit_mini --validate_only

(back to top)

Pipeline overview

┌────────────┐   ┌────────────┐   ┌────────────┐   ┌────────────┐   ┌────────────┐
│ Collection │──▶│ Conversion │──▶│  Training  │──▶│   Deploy   │──▶│    Eval    │
│ (C++, gamepad│  │  rosbags → │   │ fine-tune  │   │  real-time │   │  offline   │
│  -driven    │  │ LeRobot    │   │  VLA policy│   │  inference │   │  analysis  │
│  recording) │  │  dataset   │   │            │   │  on robot  │   │  of logs   │
└────────────┘   └────────────┘   └────────────┘   └────────────┘   └────────────┘

1. Collect demonstrations by teleoperating the robot with a gamepad (sobits_vla_rosbag_collection).
2. Convert recorded rosbags into a LeRobot dataset (sobits_vla_rosbag_conversion).
3. Train a VLA policy on the dataset (sobits_vla_training).
4. Deploy the trained policy for autonomous control (sobits_vla_deploy).
5. Evaluate the resulting episode logs offline (sobits_vla_deploy's vla_eval).

(back to top)

Namespace model

Every node advertises its own topics/services privately: ~/<channel>, which resolves to /<node_name>/<channel> bare or /<robot_name>/<node_name>/<channel> under a namespaced launch. Other nodes address it by the owner's relative name, <owner_node>/<channel> (e.g. the gamepad client's command_service defaults to sobits_vla_deploy/command or vla_rosbag_collection/command), which resolves alongside it under the same namespace. Robot I/O topics (joint states, cameras, cmd_vel — from the descriptor) stay absolute. A bare ros2 run with no namespace keeps everything under /, so a single robot works with no namespace at all; namespaced launches let multiple robots share one ROS domain.

Output roots

Each package writes its generated artifacts (rosbags, datasets, checkpoints, logs) under <pkg_src>/<artifact-dir>/, resolved by sobits_vla_common.output_root.output_root() regardless of whether you're running from a source tree, a colcon --symlink-install, or a regular install. Every <artifact-dir>/ carries a uniform *\n!.gitignore\n ignore — the directory itself is tracked, its generated contents are not. See each package's README for its specific output path.

(back to top)

Getting started

Prerequisites

System	Version
Ubuntu	22.04 (Jammy Jellyfish)
ROS	Jazzy Jalisco
Python	≥3.10

Note

If you need to install Ubuntu or ROS, please check our SOBITS Manual.

Installation

cd ~/colcon_ws/src/
git clone https://github.com/TeamSOBITS/sobits_vla_tools
cd sobits_vla_tools/
bash install.sh
cd ~/colcon_ws
rosdep update
rosdep install --from-paths src -y --ignore-src
colcon build
source install/setup.bash

Non-ROS Python dependencies (numpy, pandas, torch, lerobot) live in per-package pixi environments, not the system interpreter — see pixi.toml at the repo root and each package README's "How to run" section for the enable_gpu:=/pixi_env:= launch args that select them.

Quickstart

# 1. Record a demonstration (gamepad-driven, real or sim robot)
ros2 launch sobits_vla_rosbag_collection rosbag_collection.launch.py enable_world_reset:=false robot_name:=sobit_home

# 2. Convert the recorded rosbags into a LeRobot dataset
ros2 launch sobits_vla_rosbag_conversion rosbag_conversion.launch.py robot:=sobit_home

# 3. Train a policy on the dataset
ros2 launch sobits_vla_training sobits_vla_training.launch.py robot:=sobit_home_left_smolvla_fft steps:=30000

# 4. Deploy the trained policy
ros2 launch sobits_vla_deploy sobits_vla_deploy.launch.py deploy_config:=deploy_config_sobit_home robot_name:=sobit_home

Every argument above is a real, verified launch arg — run any file with --show-args for the full list.

(back to top)

Acknowledgments

• LeRobot — Dataset format and training framework
• ROS 2 Jazzy — Robot middleware

(back to top)