🔘 About this critique
What made the car turn? The answer begins with a distance sensor on the vehicle, passes through a dish of water on a laboratory bench and ends with a command sent back to the motor controller. Zohar and colleagues make the intermediate computation visible. Their experiment also illustrates what an independent investigator would need to account for one particular turn.
This assessment examines the published article, its supplementary methods, the physical vehicle recording and the authors’ archived software release. We checked the deposited input tables and aggregate classification results. The physical experiment and micromagnetic simulations were not repeated. Exposure describes the consequence of extending a finding beyond the conditions in which it was established.
Critique
| Examined version | Peer-reviewed article, Nature Communications 17, 10677 (10 September 2026). |
| Materials examined | Published article, supplementary methods and Video 2, and version 1.0.0 of the accompanying code and data. |
| Recommendation | Read Figure 3 alongside Figure 4c. The first measures recognition of labelled inputs. The second shows the behaviour of a vehicle on a course. |
Unit 1 — How does a distance measurement become a manoeuvre?
Article data. A single ultrasonic sensor on the physical car scans six directions. A host computer receives the distances over Bluetooth and uses them to time six mechanical excitations of a water surface. A camera records the interference pattern. After subtraction of a reference image and spatial averaging, a trained support vector machine classifies the scene. In the physical driving experiment the choice was binary. A nearby obstacle prompted a turn of approximately 30 degrees.
Reading. The wave field provides a measurable intermediate state. Its formation can be understood from the timing of the inputs and the interaction of the waves. Assigning meaning to that state requires the optical reference, the sampling procedure and the fitted classifier. A photograph of the waves alone cannot establish which reading changed the decision.
Critical conclusion. The sensing, physical transformation, classification and motor response are individually open to investigation. Explaining a particular turn requires linking observations from all four stages to the same instant.
| Claim type | Interpretability of an individual control decision |
| Exposure | High. A visible interference pattern could be mistaken for a complete explanation of the manoeuvre. |
Unit 2 — Which scenes did the classification test cover?
Article data. The study tests tasks with two, three and five output classes. The authors sampled a single water-wave image on a 50 × 50 grid, 150 ms after input excitation, with adjacent distance levels encoded 200 ms apart. For the learning curves, they report 100 test samples per evaluation and 50 repetitions with random training and test sets. Their deposited water-wave results give mean recognition rates of 97.20%, 93.46% and 87.76% at 250 training examples. All three columns reach 100% at 2,500.
| Training examples | Two classes | Three classes | Five classes |
|---|---|---|---|
| 250 | 97.20% | 93.46% | 87.76% |
| 2,500 | 100.00% | 100.00% | 100.00% |
Reading. The five-class input file contains 1,000 distinct six-distance vectors, 200 in each class. In this file, no vector places an obstacle at distance levels 1–4 on both sides at once. The accompanying generator explicitly excludes that case. These labels make it possible to check what the task asks of the classifier. A robot approaching a narrow passage may have to act precisely when both sides are occupied.
Critical conclusion. The learning curves support recognition of the defined classes under the sampling procedure used. The deposited five-class examples leave bilateral obstacles outside their coverage. The recognition percentages cannot be read as the probability of completing a journey without a collision.
| Claim type | Classification accuracy and coverage of obstacle configurations |
| Exposure | High. An error rate for labelled scenes has a different denominator from a collision rate per journey. |
Unit 3 — Can a reader reconstruct one turn without rebuilding the apparatus?
Article data. Figure 4c shows a physical car navigating a 3 m × 2 m course. Supplementary Video 2 records the run and was accelerated fiftyfold for presentation. The authors also release input generators, labelled distance vectors, aggregate water-wave scores and arrays of simulated spin-wave features. These materials let a reader check the classification rules and parts of the numerical analysis.
Reading. To reconstruct a particular turn, one would need its six distance readings, the corresponding reference and wave images, the classifier state and the command returned to the vehicle, all matched in time. We could not identify such a record for the filmed run in the archived release or the listed supplements. A trajectory supplies evidence of where the car went. Several combinations of distances and wave patterns could have produced the same turn.
Critical conclusion. A reader without the apparatus can examine the proposed mechanism, check the archived inputs and follow the recorded behaviour. The public record leaves the cause of an individual manoeuvre underdetermined. This limit concerns retrospective verification of the run. It does not imply that the apparatus is intrinsically opaque.
| Claim type | Independent reconstruction of an observed event |
| Exposure | High. Investigating an unexpected turn requires a record of the signals present at the time. |
Unit 4 — Which operating conditions have been demonstrated?
Article data. The physical car stopped periodically to take its readings and communicated with a reservoir off the vehicle. Event-triggered control was demonstrated with the virtual car. A further experiment used micromagnetic simulation to evaluate a one-micrometre magnetic disc. The supplement’s energy estimate for spin waves covers input excitation and excludes the readout and peripheral circuitry.
Reading. Each result answers a different engineering question. The physical run establishes that the sensing and command loop can close through a water reservoir. The virtual experiment shows how changes in the sensor signal can trigger computation. The simulation concerns magnetic dynamics and the information available to a readout. The 150 ms image sampling time does not measure the time from an obstacle entering the physical vehicle’s path to the response of its wheels.
Critical conclusion. A practical vehicle would need measurements of response time, collisions and whole-system energy over repeated journeys under stated operating conditions. The authors explicitly identify fabrication and readout as further steps in assessing a magnetic device.
| Claim type | Transfer from a bench experiment to an integrated vehicle |
| Exposure | High. Both the physical boundary of the energy estimate and the distinction between measured and simulated operation affect deployment decisions. |
Overall assessment
The strongest result is the closed control loop. Measured distances altered a physical wave field, a trained readout interpreted an image of that field, and the vehicle responded. By reporting how the image was formed and sampled, the authors give other laboratories specific observations against which to test the mechanism. Their archived data also make it possible to identify scene configurations that a headline recognition rate leaves out.
For a field investigation, the decisive addition would be a synchronised record of each manoeuvre. It would allow an investigator to ask whether a missed obstacle first disappeared during sensing, during wave encoding, at classification or when the command reached the vehicle. That question remains open even when the intended behaviour is visible in a successful run.
🔘 Paper page: https://www.nature.com/articles/s41467-026-77661-3
Abstract
Equipping robots with a physical, brain-inspired computing unit is a key step towards achieving true autonomy in intelligent machines; however, the practical realization of such neuromorphic hardware remains a major challenge. Because both software- and hardware-based spiking neural network approaches demand substantial training and energy consumption, an architecturally simple neuromorphic platform is required to enable scalable and energy-efficient robotics. Here, we experimentally demonstrate robotic control driven by a neuromorphic hardware based on physical wave interactions. Using a water-based experimental system that provides direct visual access to wave-network dynamics, we establish a wave-based reservoir computing framework that achieves near-perfect accuracy in robotic vehicle obstacle recognition, followed by real-time autonomous motion with obstacle avoidance controlled entirely by this wave-based hardware. Through micromagnetic simulations, we extend this concept to electrically excited spin waves in a magnetic nanodevice operating at gigahertz frequencies, outlining an approach toward solid-state robotic chips based on wave computing. These results position wave-based computation as a scalable and energy-efficient alternative to conventional digital neuromorphic architectures for next-generation intelligent machines.
Authors’ abstract reproduced from Nature Communications under the Creative Commons Attribution 4.0 International licence.
Authors
Jacob Zohar, Daniele Pinna, Gerrit van der Laan, Thorsten Hesjedal, C. K. Safeer
Liked this post? Follow this blog to get more.