A low-cost and lightweight 6 DoF bimanual arm for dynamic and contact-rich manipulation

TL;DR

Dynamic manipulation is challenged by high inertia, limited compliance, and reliance on expensive torque sensors. ARMADA addresses this by integrating low-inertia, base-mounted actuators with a 1:10 gear ratio and a 1:1 parallelogram transmission to elbow, enabling fast, compliant motion for dynamic tasks. In experiments, ARMADA achieves end-effector speeds up to 6.16 m/s, handles a 2.5 kg payload, and sustains safe contact forces, while RL-based non-prehensile policies trained in simulation transfer zero-shot to the real robot and human-motion shadowing demonstrates bimanual throwing. The work provides an open-source, low-cost platform (approximately $6k) that enables rapid prototyping and learning-driven dynamic manipulation in lab settings, potentially accelerating research in dynamic and contact-rich manipulation. ARMADA also highlights trade-offs between precision and safety, and outlines concrete paths to enhance torque density and wrist DoF for broader bimanual capabilities.

Abstract

Dynamic and contact-rich object manipulation, such as striking, snatching, or hammering, remains challenging for robotic systems due to hardware limitations. Most existing robots are constrained by high-inertia design, limited compliance, and reliance on expensive torque sensors. To address this, we introduce ARMADA (Affordable Robot for Manipulation and Dynamic Actions), a 6 degrees-of-freedom bimanual robot designed for dynamic manipulation research. ARMADA combines low-inertia, back-drivable actuators with a lightweight design, using readily available components and 3D-printed links for ease of assembly in research labs. The entire system, including both arms, is built for just $6,100. Each arm achieves speeds up to 6.16m/s, almost twice that of most collaborative robots, with a comparable payload of 2.5kg. We demonstrate ARMADA can perform dynamic manipulation like snatching, hammering, and bimanual throwing in real-world environments. We also showcase its effectiveness in reinforcement learning (RL) by training a non-prehensile manipulation policy in simulation and transferring it zero-shot to the real world, as well as human motion shadowing for dynamic bimanual object throwing. ARMADA is fully open-sourced with detailed assembly instructions, CAD models, URDFs, simulation, and learning codes. We highly recommend viewing the supplementary video at https://sites.google.com/view/im2-humanoid-arm.

Figures (11)

• Figure 1: Comparison of gearboxes in low-gear ratio actuators (b) vs. high-gear ratio actuators (c). All gear mechanisms introduce backlash and friction (a) which are difficult to model. Because low gear ratios use fewer gears, it is easier to model.
• Figure 2: Positions of the six actuators. Four heavy and strong actuators are attached to the base. Two small actuators to rotate the wrist is attached to the elbow and wrist.
• Figure 3: FEM analysis of deformation under a 10 N load applied along the (a) z-axis and (b) x-axis for configurations with all PLA and partially aluminum components. Deformation reduces from 1.69 mm to 1.58 mm (z-axis) and 2.85 mm to 2.24 mm (x-axis) with aluminum reinforcement. Deformation is exaggerated for clarity and visualization.
• Figure 4: (a) Assembly of the upper arm. Black parts are aluminum-machined, and white parts are 3D printed PLA. (b) Assembly of the custom-designed gripper. The brown component is the Dynamixel XM430 motor, the white parts are PLA parts, and the green parts are flexible TPU fingers.
• Figure 5: (a) Trajectory of the end-effector speed test. Starting from the initial joint position, ARMADA accelerates toward the terminal position. (b) A sequence of a realistic dumbbell lifting task. The robot grasps, lifts, holds for three seconds, and places down the dumbbell. (c) We use push-pull gauge to measure the maximum impact force near the point of maximum speed.