TidyVerse

TidyVerse is a research project on long-horizon household manipulation with dexterous mobile robots.
The goal is to enable robots with multi-finger hands and a mobile base to perform room-scale tidying tasks—such as arranging shoes and folding clothes—by learning from unstructured human videos.

Demos

Shoes arrangement in the garage

Slides

TidyVerse Project Slides

Documentation

For a deeper understanding of the system, see the docs/ folder:

• docs/dragonbot_control_architecture.md — full control stack architecture
• docs/teleop_testing_guide.md — step-by-step teleoperation testing guide

Robot Platform

Dragonbot v0.2

Dragonbot v0.2 is our custom dexterous mobile manipulation platform combining a holonomic base, a YAM 6-DOF arm, and an Aero Hand Open dexterous hand.

URDF models and MuJoCo scene files for visualization and simulation are available in simulation/:

• simulation/dragonbot/ — full Dragonbot scene
• simulation/yam_with_hand/ — YAM arm + hand combined scene
• simulation/assets/ — STL meshes for the base, arm, and gripper

ROS2 Full-Stack Controls

A complete ROS2-based communication and control stack is available in ros2/, covering:

• Base: SpaceMouse teleoperation (dragonbot_teleop)
• Arm: Quest 3 wrist-pose IK teleoperation (dragonbot_teleop)
• Hand: Quest 3 dexterous retargeting to 16-DOF joint commands (dragonbot_teleop)
• Hardware drivers: Aero Hand Open node and message types (aero_hand_open, aero_hand_open_msgs)
• Simulation viewers: MuJoCo passive viewers for base + arm + hand

The dexterous hand hardware and ROS2 drivers are based on TetherIA/aero-hand-open. We thank the TetherIA team for open-sourcing the Aero Hand Open platform.

GR00T-N1.6 Control Policy

We are adapting NVIDIA GR00T-N1.6 as the visuomotor control policy for Dragonbot. Our ongoing fork with Dragonbot embodiment configs, a ROS2 policy bridge, and deployment guides is available at:

👉 dragonlong/Isaac-GR00T (see third_party/Isaac-GR00T/)

Key additions in the fork:

• examples/tidyverse-hand/ — ROS2 policy bridge, FastDDS config, deployment guide
• gr00t/configs/data/embodiment_configs.py — Dragonbot embodiment definition
• dragonbot_finetune.md — step-by-step fine-tuning guide

Overview

TidyVerse focuses on three key challenges:

• Long-horizon tasks involving multiple objects and sequential subgoals
• Dexterous manipulation with multi-finger robot hands
• Human video supervision without paired human–robot demonstrations

We introduce TidyMimic++, an intent-level imitation framework that distills object-centric task structure from human videos and grounds it in robot execution.

Tasks

Example tasks in TidyVerse include:

• Shoe arrangement and alignment
• Cloth flattening and folding (TBD)
• Room-scale cleanup with navigation and manipulation (TBD)

Progress & Roadmap — v0.2 Pre-release

✅ Done

• Dragonbot v0.2 hardware platform — holonomic base + YAM 6-DOF arm + Aero Hand Open (16-DOF dexterous hand)
• Simulation models — MuJoCo URDF/XML scenes for Dragonbot, YAM+hand, and Stanford TidyBot in simulation/
• ROS2 full-stack controls — SpaceMouse base teleoperation, Quest 3 wrist-pose IK arm teleoperation, Quest 3 hand retargeting; all in ros2/
• Data pipeline — rosbag/MCAP → LeRobot dataset conversion; sample: littledragon/evan_house_split_1g_lerobot_v3
• GR00T-N1.6 integration — Dragonbot embodiment config, ROS2 policy bridge, and deployment guide in third_party/Isaac-GR00T/examples/tidyverse-hand/
• pi-0.5 fine-tuning + serving (LIBERO) — W&B logs

🔧 In progress

• GR00T-N1.6 fine-tuning on Dragonbot data — see third_party/Isaac-GR00T/dragonbot_finetune.md for instructions (coming soon)
• pi-0.5 SFT with adaptive hand grasping

🧪 Planned

• Co-training with human egocentric data (VITRA) + EgoDex
• End-to-end spatial grounding + sequential task execution

---

pi-0.5 fine-tuning details (LIBERO)

• Logs: W&B run group

License

For research use only.