The study, published in Frontiers in Neurorobotics by Pierre Klintefors, Birger Johansson, and Christian Balkenius, draws its inspiration from predictive processing, a leading theory of how the brain works. In that framework, perception and movement are unified under a single goal: minimizing the gap between predicted and observed sensory input. Movement itself is framed as a set of proprioceptive predictions—predictions about what the body will feel—rather than explicit motor commands. Sensory signals that match those predictions can be attenuated or ignored; signals that diverge demand attention. The Lund team asked a deceptively simple question: instead of treating prediction error as waste to be minimized, can a robot mine it for information?

Their answer came from Epi, an in-house two-degree-of-freedom robotic head built for developmental robotics research. The head is actuated by two Dynamixel MX-106R servomotors, one controlling pan and one controlling tilt, and carries an onboard UM7 inertial measurement unit with accelerometers and gyroscopes. Crucially, the team deliberately withheld tactile and visual feedback. The robot’s only windows onto the world were its own joint positions, its inertial readings, and the electrical current drawn by its servos—a value that rises and falls with joint load, making it a natural proxy for force. This constrained sensorium mirrors, in a rough way, the proprioceptive signals that human muscle spindles deliver, and it allowed the researchers to isolate just how much can be inferred from the body’s internal senses alone.

The first step was to give the robot an internal generative model—a learned sense of its own physics. The researchers collected training data through a process directly analogous to infant motor babbling: 2,000 randomized servo transitions, each separated by a 500-millisecond pause, with proprioceptive and inertial data sampled at 40 hertz throughout. From these recordings they trained windowed multilayer perceptrons to predict the servo current at the next time step from histories of position, distance-to-goal, inertial signals, and past current values. Each network stacked four fully connected hidden layers of 256, 128, 64, and 32 units, with LeakyReLU activations, batch normalization, L2 regularization, and dropout. Training used the Adam optimizer with a Huber loss chosen for robustness against outliers, and every configuration was trained three times with different random seeds to guard against initialization luck—54 training runs in all.

The architecture choice was itself a pragmatic engineering decision. While recurrent networks like LSTMs handle sequences elegantly, their step-by-step execution can bottleneck real-time deployment on modest hardware. A windowed MLP evaluates its entire temporal window in a single feedforward pass, which made it easy to convert to TensorFlow Lite and run on the robot’s MacBook-based control system at a strict 40-hertz inference rate. When the team compared their networks against linear ARX baselines given identical lagged inputs, every MLP variant won at every history length tested—a clear sign the networks were exploiting nonlinear structure in the sensorimotor stream rather than simply extrapolating a linear trend. One striking finding: giving the network just one step of current history (25 milliseconds) was enough, and performance stayed flat out to fifteen lags. Even more tellingly, once the model had three or more time steps of proprioceptive history, the inertial measurement unit became redundant. With enough short-term memory, the network could implicitly derive velocity and acceleration on its own—no specialized vestibular sensor required.

With the best model deployed in real time on Epi, the perturbation experiments began. The team tested five conditions: a brief manual push applied to the stationary head, a sustained force held throughout a trial, a rigid obstacle that intercepted the head during a 30-degree movement, and two unperturbed baselines—a gentle 5-degree movement and a full 30-degree sweep. Fifty trials per condition per servo yielded 500 classified trials. The manual pushes were deliberately varied in amplitude, ensuring the categories represented a range of real-world forces rather than one calibrated jab. From each two-second trial, the researchers distilled the resulting prediction error signal into twenty features: magnitude statistics such as maximum absolute error and cumulative error, temporal descriptors such as time-to-peak, half-decay time, lag-1 autocorrelation, and a temporal center of mass, and morphological shape coefficients drawn from the discrete cosine transform of the normalized signal.

The resulting profiles were vividly distinct. A brief push produced a sharp spike in prediction error that stabilized within about 400 milliseconds. Sustained force left a long, persistent tail in the error signal. An obstacle collision drove the error amplitude far above that of an identical unobstructed movement. The two unperturbed movements resembled each other, with larger amplitude peaks for the wider 30-degree sweep. A random forest classifier, trained under rigorous nested cross-validation with tilt and pan data kept in separate trials to prevent information leakage, achieved 98.5 percent overall accuracy across the five conditions, with a 95 percent bootstrap confidence interval spanning 97.4 to 99.6 percent. A permutation test confirmed the result was no statistical fluke: across 100 label-shuffled repetitions, none came close to the observed accuracy, yielding an empirical p-value of 0.0099.

Perhaps the most philosophically interesting result concerned what the classifier actually needed. Magnitude features alone reached 95.9 percent accuracy. But the amplitude-normalized temporal and shape features—those carrying pure temporal structure, stripped of how big the errors were—performed even better at 97.6 percent. The two most discriminative features were the temporal center of mass of the error signal and the mean error in the first of five phase bins, together accounting for over a quarter of the classifier’s importance weight. And when the researchers classified the raw servo current directly, bypassing the prediction error entirely, they got exactly the same 98.5 percent. In other words, the prediction error carried just as much diagnostic information as the raw signal it was derived from. The mismatch between expectation and reality turned out to be as informative as reality itself.

Why does this matter beyond a lab in Lund? The team suggests that error profiles could directly drive behavior in future systems: a sharp spike characteristic of a push might trigger reorientation or stabilization, sustained force might trigger compliance so the robot yields gracefully, obstacle contact might trigger an immediate halt, while expected movement-related errors would be attenuated and ignored. This resonates with sensory attenuation in biology, where self-generated sensations evoke weaker neural responses than externally caused ones—why you cannot tickle yourself, in essence. It also echoes classical control theory: the researchers blended predicted and actual current at a fixed 80-to-20 ratio, a design choice analogous to the uncertainty-weighted gain of a Kalman filter, keeping the error signal stable without drift.

The authors are candid that this is a proof of principle in a deliberately constrained setting. Perturbations came from a single operator pushing in a consistent direction, obstacle interactions occurred only during fixed 30-degree movements, and the platform has just two degrees of freedom. Scaling to robots with many coupled joints—where one limb’s load depends on the state of others—remains an open challenge, as does handling perturbation directions, magnitudes, and movement contexts beyond those tested. Hand-crafted features may also need to give way to learned temporal embeddings for messier interactions. Yet the conceptual payoff stands: imperfect prediction need not be a liability. The very structure of a robot’s surprise—its onset, its peak, its decay, its persistence—encodes what the world just did to it. In the predictive processing view the authors embrace, error is not merely a signal to silence but a resource to read, and machines that learn to read it may one day sense their world through their bodies alone, much as we do.

Subject of Research: Temporal profiles of load-related prediction errors for classifying physical perturbations on an embodied robot

Article Title: Temporal signatures of prediction error reveal external forces on an embodied robot

Article References: Klintefors, P., Johansson, B., & Balkenius, C. (2026). Temporal signatures of prediction error reveal external forces on an embodied robot. Frontiers in Neurorobotics, 20, Article 1918275. https://doi.org/10.3389/fnbot.2026.1918275

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1918275

Keywords: cognitive robotics, predictive processing, prediction error, proprioception, perturbation classification, embodied robotics, servo current, motor babbling, active inference, neurorobotics, sensory attenuation, machine learning