Low-Cost Mobile Bimanual Manipulation for Long-Horizon Toilet Cleaning
TL;DR A mobile bimanual robot costing about €3,200, built from two LeRobot SO-101 arms on a TurtleBot 4, used to compare open VLA policies on a five-stage toilet-cleaning task. Fine-tuned on 250 real demonstrations, π0 does best with 60% strict success (90% with one human intervention); π0.5, SmolVLA, GR00T and the memory- and progress-augmented variants all do worse.
What they do
They build a low-cost, reproducible mobile bimanual robot and use it to compare open vision-language-action (VLA) models on the same robot, data and task. The task is long-horizon toilet cleaning in five stages: grab a sponge and drive to the stain, scrub with the rough side, flip the sponge with both arms, wipe with the soft side, then put the sponge away and return.
How it works
- Hardware. Two 5-DoF SO-101 arms sit on a 3D-printed triangular base on a redesigned TurtleBot 4 top cover, with a mast-mounted top camera and a wrist camera on each arm. The platform costs about €3200, of which the arm extension is about €500.
- Software. A LeRobot–ROS 2 stack runs policy inference on a workstation, while an onboard Raspberry Pi 4 handles low-level control. Two operators teleoperate, one with leader arms and one driving the base by keyboard.
- Data and policies. 250 demonstrations at 30 Hz, 200 teleoperated and 50 DAgger-style corrections. They fine-tune π0, π0.5, SmolVLA and GR00T N1.7, plus π0 with SARM task-progress inputs and π0.5 with short-term MEM memory.
Results
Each policy gets 20 rollouts (GR00T 5), with toilet and table orientation, sponge position and stain randomized:
- π0: 60% strict, 70% relaxed, and 90% with at most one human repositioning. It mostly failed by getting stuck under the toilet seat or not moving on to the next subtask.
- π0 with SARM: 40% / 55% / 65%, worse than π0. π0.5 reached 25% strict success; π0.5 with MEM, 10%.
- SmolVLA and GR00T: 0%. In most rollouts SmolVLA never reached the sponge; GR00T could not pick it up.
Limitations
One task in one setting, with lighting, background, toilet and tools held fixed. The 5-DoF arms limit reach and dexterity, and the robot relies on external power and an off-board workstation. The released π0.5 includes only the flow-matching action head, and why the augmented variants did worse was not isolated by ablations.
Figures
Abstract
Vision-Language-Action (VLA) models have shown promising capabilities for robot manipulation, but their performance on accessible mobile bimanual platforms remains underexplored. We present a low-cost and reproducible mobile bimanual platform integrating two LeRobot SO-101 manipulators with a TurtleBot 4 and a LeRobot-ROS 2 software stack. Using long-horizon toilet cleaning as a case study, we collect 250 real-world demonstrations and evaluate $π_0$, $π_{0.5}$, SmolVLA, and GR00T under the same embodiment and task conditions. We further evaluate variants using SARM-derived task-progress signals and short-term MEM observation history. Among the evaluated models, $π_0$ performs best, achieving 60% strict and 70% relaxed autonomous success, increasing to 90% with limited human intervention. In contrast, the evaluated progress- and memory-augmented variants do not improve their corresponding baselines. Our results demonstrate the feasibility of long-horizon mobile bimanual manipulation on accessible hardware while highlighting remaining challenges in reliable VLA deployment. Project website: https://so101-turtlebot4.github.io/mobile-bimanual-cleaning/
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.