The dream of "fleet learning"—where robots share data to master complex tasks—has been hampered by a hardware divide. On one side, we have inexpensive, position-controlled arms that are stiff and hard to simulate. On the other, we have industrial-grade, torque-controlled manipulators that cost as much as a luxury car.
TinyARM-X bridges this gap. It is a new open-source framework for building desktop-scale, torque-controllable robotic manipulators for under $350. By combining clever actuator modeling with a modular hardware design, the authors have created a platform that makes sophisticated force-aware manipulation accessible to researchers and hobbyists alike.
TL;DR
TinyARM-X is an open-source, modular robot framework that brings high-fidelity torque control to the sub-$350 price point. By leveraging inexpensive servos and advanced dynamic modeling, it enables force-sensitive tasks, haptic feedback, and seamless sim-to-real transfer on a desktop scale.
What Problem Does This Paper Address?
Most low-cost robots (like the popular SO-101) rely on rigid position control. While great for simple pick-and-place, these robots are "hard" to backdrive due to high gear ratios. This makes them dangerous during unexpected collisions, difficult to use for contact-rich tasks (like whisking eggs or plugging in a USB), and creates a massive "sim-to-real" gap because their physics are hard to model accurately.
True torque-controlled robots solve these issues by allowing the robot to be compliant and "feel" its environment. However, these systems typically require expensive joint-torque sensors and high-end actuators. TinyARM-X aims to provide this "transparency" using hobby-grade Dynamixel XL330 servos, which are traditionally limited to position control.
Methods
The authors achieved "industrial" functionality through a combination of modular hardware and a sophisticated software stack:
-
Common Actuation Units (CAUs): The robot is built from standardized units that can be 3D-printed and assembled with just two screws per link. This allows a user to reconfigure a robot from 3 to 7 Degrees of Freedom (DoF) in under five minutes. Overview of the TinyARM-X framework showing modularity and the software stack.
-
Non-Linear Actuator Modeling: Cheap servos don't have a linear relationship between input current and output torque. The authors implemented a square-root model ($\tau \propto \sqrt{i}$) to accurately map current commands to actual forces.
- Alpha-Beta State Estimation: To get clean velocity data from noisy, quantized position sensors, they used an alpha-beta filter. This is crucial for calculating the damping needed for stable impedance control.
- Dynamic Compensation: Using the Pinocchio library, the system calculates gravity, Coriolis, and friction forces in real-time. By "canceling out" these forces, the robot becomes "transparent"—it stays exactly where you put it and moves with the slightest touch.
- Cartesian Impedance Control (CIC): Instead of controlling joint angles, CIC allows the user to define the "stiffness" of the end-effector in 3D space, making teleoperation much more intuitive.
Datasets
The paper does not introduce a large-scale training dataset. Instead, it focuses on mechanical benchmarking and teleoperation performance. The authors evaluated the system using:
-
Backdrive Benchmarks: Measuring the force required to move the arm when powered vs. unpowered using an industrial Flexiv Rizon robot as a reference. The TinyARM-7 coupled to an industrial Flexiv Rizon 10s for force measurement.
-
MR Teleoperation Trials: A user wearing a Meta Quest 3 headset attempted to apply specific forces (100g, 150g, 200g) to a scale without seeing the scale's display.
Key Insights and Ideas
- Software-Defined Embodiment: The software automatically scans for servo IDs on startup. It detects how many joints are connected and automatically loads the corresponding URDF model for dynamics calculations.
- Redundancy as a Tool: In the 7-DoF version, the authors use "null-space projection." This allows the robot to perform auxiliary tasks (like moving its elbow to avoid an obstacle) without changing the force or position at the gripper.
- The Power of Transparency: By modeling friction and gravity, a $350 arm can function as a high-end haptic input device, allowing a user to "feel" virtual objects in a MuJoCo simulation.
Key Results
- Cost & Accessibility: The total Bill of Materials (BOM) is approximately $346.65.
- Rapid Reconfiguration: Swapping a joint takes an average of 259 seconds.
- Mechanical Performance: The TinyARM-7 has a reach of 210mm and a max speed of 0.77 m/s. It achieves a backdrive-breakaway force of just 1.4 N.
- Force Precision: In Mixed Reality (MR) tests, using Cartesian Impedance Control resulted in less than 10% error when applying specific forces, whereas standard joint-space control led to errors near 50%. A user in an MR headset teleoperating the arm to apply exactly 100g of force to a scale.
Limitations and Future Work
The primary limitation is payload. At full extension, the TinyARM-7 can only lift about 33.5g before the shoulder servos saturate. While this is enough for many desktop tasks (like handling pens or small components), it limits the "strength" of haptic feedback.
The authors suggest that future versions could use doubly-actuated joints (two servos per joint) in the shoulder and elbow to significantly increase payload and reach. They also aim to automate the generation of URDF models for any arbitrary user-created linkage, further lowering the barrier to custom robot design.
Summary updated