Excited to share that our work NeuralActuator: Neural Actuation Modeling for Robot Dynamics and External Force Perception has been accepted to #RSS2026! Your robot — even a low-cost one — can feel external forces without torque or tactile sensors. TL;DR: NeuralActuator is a neural actuator model that jointly predicts... 1️⃣torque to capture the nonlinear and time-varying current–to–torque relationship of low-cost servos, 2️⃣external contact forces (and force detection gates) for sensorless force perception, 3️⃣and motor conditions that indicate each motor’s operating regime. Here is a fast-forward video clip ⬇️ We are also covering more robots like LeRobot-S101 and Franka Panda. More details coming soon.show more

Robot dynamics, [ M(q)\ddot q + C(q,\dot q)\dot q + g(q) = \tau + \tau_{\mathrm{ext}}, ] suggest that motion, motor torques, and external interactions are tightly coupled. In principle, this coupling enables external forces to be inferred from the robot’s state and actuation. But on low-cost robot platforms, exploiting this relationship is hard: the linear current–torque model often breaks down during target tracking due to friction, hysteresis, backlash, and thermal effects. We propose NeuralActuator, a learning-based method with differentiable simulation for actuator modeling, enabling accurate dynamics modeling, sensorless force estimation, and motor condition estimation.

Sincere thanks to all our collaborators: John U. Onyemelukwe, Hangxing Zhang, Heng Zhang, Minghao Guo @GuoMh14, Yunsheng Tian @yunshengtian, Michal Piotr Lipiec, Joshua Jacob, Chao Liu, Peter Yichen Chen @peterchencyc , Yuri Ivanov, and Wojciech Matusik. @wojmatusik 🙌

Is there a link to the paper?

Yes! We are working on the camera-ready one. Will release it very soon.

coool

@RemiCadene Awesome

@NepYope Wait… dynamixel XM motors ARE linear wrt current and torque output though???

Yes, at the motor level, torque is usually modeled as \tau = K_t I, where K_t​ is the torque constant and I is the motor current. This seems to be supported by the ROBOTIS performance graph as well: However, in practice, for these relatively low-cost servo actuators, the relationship between current and actual output torque during operation is not very cleanly linear. It can be quite noisy and nonlinear due to gearbox friction, backlash, saturation, current-control behavior, voltage variation, and other actuator-level effects. The figure below shows the simulation-based surrogate torque (derived from a differentiable simulation) and the measured current for each joint. Note that it is very difficult to directly measure ground-truth torque on this kind of robotic arm. From the plots, we can see that the surrogate torque and current do not follow a simple linear relationship very well.

@NepYope That is true, but I’ve literally measured different torque outputs from those actuators when they’re bolted down to tables. When the current controller is active, it’s linear. Like R^2=.99 linear. For lower cost servos, it’s a little curved (it’s a sqrt function), also R^2=.99

@NepYope Yes, but the measurement is taken while the entire robotic arm is moving and executing the task. In this setting, all motors are part of the articulated structure of the robotic arm, which is different from measuring each motor directly in isolation.