cs.ROMar 10, 2026
SaveCutting the Cord: System Architecture for Low-Cost, GPU-Accelerated Bimanual Mobile Manipulation
Abstract
We present a bimanual mobile manipulator built on the open-source XLeRobot with integrated onboard compute for less than $1300. Key contributions include: (1) optimized mechanical design maximizing stiffness-to-weight ratio, (2) a Tri-Bus power topology isolating compute from motor-induced voltage transients, and (3) embedded autonomy using NVIDIA Jetson Orin Nano for untethered operation. The platform enables teleoperation, autonomous SLAM navigation, and vision-based manipulation without external dependencies, providing a low-cost alternative for research and education in robotics and robot learning.
Figures & tables
Fig. 1: The evolved XLeRobot platform a low-cost bimanual platform for education and research assisting an industrial robot by removing screws from an electric vehicle battery.
Fig. 2: Diverse Bimanual Manipulation Tasks. Demonstrations include pencil insertion into an electric pencil sharpener, picking up a YCB drill [ 13 ] , bimanual container opening, toaster use, screw removal from car-battery disassembly, whiteboard wiping, peg-in-hole insertion, coffee filter removal, block organization, and drawer opening.
| Platform | Compute | GPU | Untethered | DoF | Payload | Est. Cost |
|---|---|---|---|---|---|---|
| AhaRobot [ 14 ] | Mini-ITX (i5) | RTX 4060 | Yes | 16 | 1.5 kg/arm | $2k |
| Cone-E [ 5 , 15 ] | Intel NUC | — | Yes | 16 | 1.5 kg/arm | $12k |
| Mobile ALOHA [ 8 ] | Laptop (onboard) | RTX 3070 Ti | Yes | 16 | 0.75 kg/arm | $32k |
| TB3 Waffle Pi + arms [ 16 , 17 ] | Raspberry Pi 4 | — | Yes | 12 | 0.5 kg/arm | $3k |
| XLeRobot [ 11 ] | External PC | — | No | 15 | 0.8 kg | $700–1k |
| Ours | Jetson Orin | Ampere | Yes | 17 | 1.0 kg/arm | $1.2k |
TABLE I: Comparison of Dual-Arm Mobile Manipulators
| Subsystem | Component | Qty | Cost (USD) |
| Compute | NVIDIA Jetson Orin Nano Super [ 18 ] | 1 | $249.00 |
| Compute | microSD Card | 1 | $11.23 |
| Peripherals | Anker 4-in-1 USB-C Hub | 1 | $14.99 |
| Actuation | Feetech STS-3215 Servos (12V) [ 21 ] | 17 | $271.83 |
| Actuation | Servo Wire: 50’ 3-Color | 1 | $15.97 |
| Actuation | Wonrabai Serial Bus Servo Driver Board | 2 | $21.10 |
TABLE II: System Bill of Materials (BOM)
Fig. 3: Structural composition and kinematics: (left) Robotic arm with 4-wall (black)/2-wall topology (cyan), and (right) cantilever experimental setup.
Fig. 4: Power System Architecture. (a) The logical power schematic illustrates the distribution from the Anker SOLIX C300 to the distinct motor buses and the NVIDIA Jetson controller via high-output USB-C and car outlet interfaces. (b) The hardware integration diagram shows the physical routing of power lines within the mobile manipulator chassis, supporting the Intel RealSense D435 and dual arm configurations.
| Bus Group | Torque | Accel | Load | Limit | PDU |
| ( ) * | ( ) | ( ) | (Fuse) | Constraint | |
| Bus A (Wheels/Neck) | 650 | 20 (Med) | 9.84 A | 10 A | USB-C2 Output |
| Bus B (Arms) | 450 | 40 (Soft) | 4.90 A | 5 A | DC Car Outlet |
| Jetson Orin | N/A | N/A | 2.1 A | 5 A | USB-C3 Output |
TABLE III: Firmware Saturation Settings
Fig. 5: Perception-driven manipulation pipeline. (a) Color-based segmentation and centroid detection in the camera optical frame. (b) 3D mapping and representation of the environment. (c) Final grasp execution of the target object following inverse kinematics planning.
Fig. 6: SLAM Navigation Representation. (a) “Ground truth” top down view of the laboratory (Scaniverse) and (b) corresponding 2D occupancy grid for autonomous navigation.
Fig. 7: Structural Performance Characterization. Comparison of link topologies across rigidity (top) and ultimate load capacity (bottom, ).
Fig. 8: Grasp Performance and Reliability. (a-c) Compliant mechanical adaptation. (d) Experimental results demonstrating a 98.7% success rate.
| Model | (ms) | ||||
|---|---|---|---|---|---|
| ACT | 100 | 50 | - | 36.0 0.9 | 27.8 |
| Diffusion | 20 | 10 | 10 | 539.6 0.1 | 1.8 |
| SmolVLA | 20 | 10 | 10 | 713.8 9.3 | 1.4 |
TABLE IV: Inference Benchmarks on Jetson Orin Nano (MAXN SUPER Mode)
Fig. 9: VR Teleoperation Setup: (Left) The operator interacting with the system; (Right) a detailed view of the ManipulationNet peg-in-hole board.
| Modality | Intuitiveness | Ergonomics | Time (s) |
|---|---|---|---|
| VR Controller * | 2.67 | 2.33 | 48.22 |
| Hand Tracking * | 1.33 | 1.67 | 68.30 |
| Xbox Controller Baseline | 1.50 | 3.67 | 248.73 |
TABLE V: Teleoperation Performance on High-Precision Peg-in-Hole Task ( total trials)
More in this topic
Robotics
Robotics studies the design and control of machines that perceive, move, and act in physical environments.
See all 751 papers