TidyBot++ Holonomic Mobile Manipulator

• TidyBot++ is an open-source holonomic mobile manipulator designed for household research, enabling efficient imitation learning via human teleoperation.
• It integrates a kinematically agile base built with modified FRC MK4 swerve modules and modular hardware components for versatile manipulation tasks.
• Experimental evaluations show significant performance gains in maneuverability and task success when using the holonomic mode compared to pseudo-differential approaches.

TidyBot++ is an open-source, holonomic mobile manipulator platform engineered for scalable research in real-world household mobile manipulation and robot learning. Building on principles of kinematically agile base design and hardware modularity, TidyBot++ is designed to facilitate imitation learning via efficient, robust human teleoperation and provides a hardware/software ecosystem that supports a wide spectrum of manipulation tasks (Wu et al., 2024).

1. Mechanical Architecture and Kinematics

At the core of TidyBot++ is a holonomic base constructed from four powered FRC MK4 swerve modules, each with dual-motor actuation (steering and wheel) and integrated encoders on a CAN bus. Uniquely, each caster module is modified to introduce an offset between the steering axis and wheel axle—specifically 14 mm longitudinally and several millimeters laterally—such that the trailing caster aligns with the velocity vector, conferring true holonomic drive capability. This enables any planar twist $(v_x, v_y, \omega)$ of the base to be achieved instantaneously, eliminating nonholonomic constraints inherent in differential or omnidirectional wheel arrangements.

The kinematic model, adapted from Holmberg’s Powered Caster Vehicle (PCV) theory, forms the mathematical foundation:

• Body-frame twist: $V = [v_x, v_y, \omega]^T$
• For module $i$ at $(h_i, \beta_i)$:
  • Steering angle $\phi_i$
  • Wheel rotation $\rho_i$
  • Inverse kinematics:
  • $v_{x_i} = v_x - \omega h_i \sin \beta_i$
  • $v_{y_i} = v_y + \omega h_i \cos \beta_i$
  • $\phi_i^* = \text{atan2}(v_{y_i}, v_{x_i})$
  • $$\dot{\rho}_i^* = \frac{1}{r}[v_{x_i} \cos \phi_i^* + v_{y_i} \sin \phi_i^*]$$
  • Forward kinematics:
  • $V = [v_x, v_y, \omega]^T$0
  • $V = [v_x, v_y, \omega]^T$1 is formed from the configuration of all four modules.

This kinematic assembly enables fully independent control of the base’s planar degrees of freedom (DOF), supporting arbitrary translations and rotations without sequential turning or alignment. Significant advantages are realized in maneuverability, including pure lateral motions and in-place rotation, with performance gains for manipulation tasks that demand tight trajectory alignment or workspace access (Wu et al., 2024).

2. Hardware Components and Open-Source Design

The primary structural elements include an aluminum T-slot extrusion frame, laser-cut acrylic plates, and a mounting system for arbitrary manipulators. The system is powered by a 12 V SLA battery for motors and a 768 Wh portable power station for computation and peripherals.

Key hardware features:

• Kinova Gen3 7-DOF arm as default manipulator (frame supports arbitrary and dual-arm configurations)
• Intel NUC mini-PC for onboard computation
• USB-to-CAN interface for real-time swerve module command/feedback
• Power distribution for both DC (motors) and AC (compute/manip) devices

Modularity is prioritized: mounting plates and frame layouts are reconfigurable, enabling payload variation and retrofitting for other arms or sensors. Complete 3D CAD files, mechanical drawings, bill of materials, and assembly instructions are openly available (Wu et al., 2024).

| Implementation Aspect | Specification/Resource | |------------------------|------------------------------------------------------------| | Total hardware cost | \$5–6k USD (holonomic base including compute) | | Footprint | 50×54 cm | | Payload capacity | 60 kg | | Runtime | 8 h (on battery) | | CAD/assembly access | http://tidybot2.github.io |

3. Software, Teleoperation, and Control Stack

TidyBot++’s control stack is centered on ROS and features a real-time low-level base controller, a high-level interface for user teleoperation, and full data logging for imitation learning.

Teleoperation

• A mobile-phone interface, implemented using WebXR, allows the operator to transmit 6-DOF pose streams from any modern smartphone.
• The interface toggles between base and arm control, mapping either phone pose to $V = [v_x, v_y, \omega]^T$2 base commands or to 6-DOF end-effector goals.
• Communication occurs via WebSockets, and the entire demo can be acquired in real time.

Low-Level Control

• Desired base twists $V = [v_x, v_y, \omega]^T$3 are interpreted by the ROS base node, implementing the above kinematics to command module steer positions and wheel velocities.
• Odometry is computed using the forward kinematic Jacobian $V = [v_x, v_y, \omega]^T$4, with pose broadcast at $V = [v_x, v_y, \omega]^T$550 Hz.
• Arm control is integrated via ROS, enabling coordinated, real-time mobile manipulation.

Data Acquisition

Demonstrations are stored as time-aligned sequences containing:

• RGB-D imagery
• Arm and base joint states
• End-effector pose
• Teleoperation commands

This structure supports direct playback/retraining for imitation learning (Wu et al., 2024).

4. Imitation Learning and Experimental Evaluation

TidyBot++ supports practical, scalable data collection for policy learning in mobile manipulation:

• Demonstration data sets for tasks such as “open fridge,” “wipe countertop,” “load dishwasher,” “take out trash,” “load laundry,” and “water plant” are collected with only 50–100 demonstrations per task.
• Diffusion policies (as in [Chi et al. 2023]) are trained for 500 epochs and evaluated in closed loop.

| Task | Demos | Success Rate (10 Trials) | |-------------------|-------|-------------------------| | Open fridge | 100 | 10/10 | | Wipe countertop | 50 | 9/10 | | Load dishwasher | 50 | 7/10 | | Take out trash | 50 | 10/10 | | Load laundry | 50 | 7/10 | | Water plant | 50 | 6/10 |

Base odometry drift is $V = [v_x, v_y, \omega]^T$61 cm per meter and $V = [v_x, v_y, \omega]^T$71^\circV=[vx,vy,ω]TV = [v_x, v_y, \omega]^T8^\circ</sup></sup>rotation.Successfulpoliciesareachievableevenwithlimiteddata,anddatacollectionfor50episodescanbeaccomplishedwithin1–2hoursfornon−trivialtasks(<ahref="/papers/2412.10447"title=""rel="nofollow"data−turbo="false"class="assistant−link"x−datax−tooltip.raw="">Wuetal.,2024</a>).</p><h2class=′paper−heading′id=′analysis−of−holonomic−vs−differential−drive−bases′>5.AnalysisofHolonomicvsDifferential−DriveBases</h2><p>Adirectexperimentalcomparisonispresentedforthe“wipecountertop”task:</p><ul><li>Holonomicmode:supportsarbitrary</sup></sup> rotation. Successful policies are achievable even with limited data, and data collection for 50 episodes can be accomplished within 1–2 hours for non-trivial tasks (<a href="/papers/2412.10447" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Wu et al., 2024</a>).</p> <h2 class='paper-heading' id='analysis-of-holonomic-vs-differential-drive-bases'>5. Analysis of Holonomic vs Differential-Drive Bases</h2> <p>A direct experimental comparison is presented for the “wipe countertop” task:</p> <ul> <li>Holonomic mode: supports arbitrary V = [v_x, v_y, \omega]^T$9.

Pseudo-differential mode: $i$0 constraint, only $i$1 allowed.

Performance differences:

| Metric | Holonomic | Pseudo-Differential | |-------------------------|-----------|--------------------| | Mean path length (m) | 2.03 | 4.03 | | Mean duration (s) | 27.4 | 65.2 | | Policy success (10 tr.) | 9 | 4 |

The holonomic mode enables efficient lateral sweeps and smoother trajectories both for human operators and learned policies. Pseudo-differential bases produce longer, more complex paths, and fail to capture the nuanced spatial execution required for high-reliability mobile manipulation tasks (Wu et al., 2024).

6. Applications, Limitations, and Future Work

TidyBot++ is intended as a general platform for:

• Scalable demonstration collection and policy imitation learning for complex manipulation
• Research into holonomic base trajectory planning and model-based mobile manipulation
• Benchmarking data-efficient diffusion models for high-DOF control

Limitations include high friction in steering due to high gear ratios and minimal axial offset, which impedes smooth backdriving and kinesthetic teaching. Current design choices prioritize cost efficiency via COTS components, at some expense to compliance and manual teachability.

Proposed future improvements:

• Custom caster modules with reduced gearing and greater offsets to reduce friction and enable easier backdriving
• Integration of torque sensors in steer axes for improved force feedback
• Addition of on-board vision for fully closed-loop policy execution
• Expansion to dual-arm or variable-payload chassis with modular frame extensions

All design files, component lists, software repositories, and user guides are available at http://tidybot2.github.io for open community extension and reproducibility (Wu et al., 2024).