innate-inc/innate-os

Purpose and Runtime Entry Points

Innate OS is a ROS 2 Humble-based operating system for MARS robots, combining physical control, autonomous multimodal agents, skills, training, cloud integration, a browser control surface, and a MuJoCo digital twin. Primary operator entry points are the robot web interface, innate CLI, ROS launch files, and simulator CLI:

• innate-sim

• sim/launcher/main.py

• scripts/launch_ros_in_tmux.sh

• ros2_ws/src/mars_bot/mars_bringup/launch/mars_bringup.launch.py

ROS Workspace and Shared Contracts

ros2_ws is the runtime boundary: independently built ROS packages communicate through topics, services, and actions. Message packages define the control contracts between brain, manipulation, cloud, and robot components, isolating callers from implementation details.

• ros2_ws/src/brain/brain_messages/action/ExecuteSkill.action

• ros2_ws/src/brain/brain_messages/srv/ReloadSkillsAgents.srv

• ros2_ws/src/cloud/innate_cloud_msgs/action/NavigateInstruction.action

• ros2_ws/src/mars_bot/mars_msgs/srv/GotoJSTrajectory.srv

Robot Hardware, Perception, and Motion

The MARS platform layer brings up motors, base, IMU, LiDAR, battery, cameras, arm/head servos, and TF. Control accepts teleoperation or navigation velocity commands; arm services and KDL-style IK provide manipulation primitives. Camera pipelines publish stereo/depth data and can stream WebRTC remotely.

• ros2_ws/src/mars_bot/mars_bringup/mars_bringup/bringup.py

• ros2_ws/src/mars_bot/mars_control/mars_control/app.cpp

• ros2_ws/src/mars_bot/mars_arm/mars_arm/arm_node.cpp

• ros2_ws/src/mars_bot/mars_cam/mars_cam/stereo_depth_estimator.cpp

• ros2_ws/src/mars_bot/mars_cam/mars_cam/webrtc_streamer.cpp

Navigation and Manipulation Execution

Navigation wraps Nav2, mapping/localization modes, costmaps, behavior trees, command multiplexing, and custom goal/footprint behavior. Manipulation records demonstrations, manages tasks, replays episodes, and runs learned ACT/TensorRT policies. These are the primary physical-action backends invoked by skills and agents.

• ros2_ws/src/mars_bot/mars_nav/mars_nav/mode_manager.py

• ros2_ws/src/mars_bot/mars_nav/mars_nav/navigate_to_pose_router.py

• ros2_ws/src/brain/manipulation/manipulation/manipulation_server.py

• ros2_ws/src/brain/manipulation/src/recorder_node.cpp

• ros2_ws/src/brain/manipulation/manipulation/act_trt.py

Brain, Agents, Skills, Inputs, and Memory

brain_client is the autonomy orchestrator. It loads workspace-defined agents, skill packs, and input devices; gathers robot/perception state and memory; runs a model-facing agent loop; and invokes or interrupts skills through ROS-facing servers. Skills are Python extensions with injected mobility/manipulation/perception interfaces, while agents constrain allowed skills, inputs, and prompts.

• ros2_ws/src/brain/brain_client/brain_client/nodes/brain_client_node.py

• ros2_ws/src/brain/brain_client/brain_client/brain/loop.py

• ros2_ws/src/brain/brain_client/brain_client/skills/runner.py

• ros2_ws/src/brain/brain_client/brain_client/inputs/manager.py

• workspace/innate_agents/basic_agent.py

• workspace/innate_skills/navigate_to_position.py

Cloud, Training, and Observability Boundaries

Cloud packages authenticate and proxy external model/TTS services, upload logs, transfer episodes, manage training jobs, and deploy resulting skills. The brain includes Gemini and vision integration; proxy adapters expose OpenAI and Cartesia. These integrations remain clients around ROS nodes rather than replacing local robot control.

• ros2_ws/src/cloud/clients/proxy-client/innate_proxy/client.py

• ros2_ws/src/cloud/innate_training_node/innate_training_node/node.py

• ros2_ws/src/cloud/innate_logger/innate_logger/logger_node.py

• ros2_ws/src/brain/brain_client/innate/gemini.py

User Interfaces, Simulation, and Deployment

The static webapp uses rosbridge, WebRTC, and HTTPS proxy routes to operate the same ROS services/actions as mobile clients. The simulator runs the real ROS stack with simulated hardware adapters, Dockerized provisioning, a Python launcher, and a TypeScript 3D viewer connected through rosbridge. Physical deployment is systemd-managed with DDS/Zenoh, Bluetooth, networking, and device configuration.

• webapp/js/rosClient.js

• webapp/proxy/https_server.py

• sim/launcher/runtime.py

• ros2_ws/src/mars_bot/mars_sim_driver/mars_sim_driver/node.py

• sim/viewer/src/physics/rosbridgeController.ts

• config/systemd/ros-app.service

• config/systemd/zenoh-router.service