Winkler, Alexander, Mana, Neeraj Naduvath, Gordon, David, Andert, Jakob · Sep 27, 2026
arXiv eess.SY authors report faster cloned control for a hydrogen-diesel engine, with limits beyond training range
Markdown · JSON · Original source
Alexander Winkler, Neeraj Naduvath Mana, David Gordon and Jakob Andert report training a feedforward with behaviour cloning (BC)—learning to imitate an expert controller—to copy Model Predictive Control (MPC) for a hydrogen-diesel dual-fuel engine. MPC adjusts controls through online optimisation, but the authors say that computation can limit its use on low-cost embedded hardware. Their controller learned from 86,000 engine cycles recorded on a modified Cummins engine, then was experimentally tested on unseen fast-transient load steps. The authors report that both versions tracked those steps with normalised root mean square error (NRMSE) values close to the MPC expert’s, while mean NOx and particulate matter emissions were no higher than the expert’s. On a Raspberry Pi 400, inference took 2 ms or less, compared with up to 7 ms for the expert; an open-loop profile on an ESP32 also fit within the stated cycle window. These results suggest a way to reduce control computation in this engine setup, but do not establish performance across other engines or operating conditions. The authors also report a crucial boundary: beyond the training range, the cloned policy saturates its controls and exceeds the pressure-rise-rate limit.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
eess.SY · Article
Source details
Original source date: Sep 27, 2026 (2026-09-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T08:01:53.686Z. First observed: 2026-09-29T06:09:03.131Z. Retrieved: 2026-10-29T19:03:22.641Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported training data came from one modified Cummins QSB 4.5-litre hydrogen dual-fuel engine at 1500 min-1; the evidence does not establish performance on other engines. The authors report that the cloned policy exceeds the pressure-rise-rate limit beyond the training range. The ESP32 result is described as open-loop profiling, not an experimental closed-loop test on that microcontroller.
Open question
How wide an operating range can the cloned policy handle before it exceeds the pressure-rise-rate limit?
Source-stated
The authors trained a feedforward deep neural network by behaviour cloning to imitate an MPC controller, using 86,000 engine cycles recorded on a modified hydrogen dual-fuel engine.
This paper trains a feedforward deep neural network (DNN) by behaviour cloning (BC) to imitate an MPC expert, using 86,000 engine cycles of demonstration data collected at 1500 min-1 on a modified Cummins QSB 4.5-litre hydrogen dual-fuel engine.
Source-stated
In experimental tests on unseen fast-transient load steps, both cloned variants had NRMSE values close to the MPC expert's, and mean NOx and particulate matter emissions were at or below the expert's.
Both track unseen fast-transient load steps (3-8 bar indicated mean effective pressure, IMEP) with normalised root mean square error (NRMSE) values of 7.80% and 9.03% against the expert's 8.01%, while keeping mean NOx and particulate matter emissions at or below those of the expert.
Source-stated
The authors report faster inference than the MPC expert on a Raspberry Pi 400 and an open-loop inference time within the stated cycle window on an ESP32.
Inference takes 2 ms or less on a Raspberry Pi 400 (ARM Cortex-A72 at 2.2 GHz), including 1 ms communication latency, compared to up to 7 ms for the MPC expert, a 3.5x speedup.
Source-stated
Beyond the training range, the cloned policy saturates its controls and exceeds the pressure-rise-rate limit.
Beyond the training range the cloned policy saturates its controls but exceeds the pressure-rise-rate limit.
Source and contributor context
Winkler, Alexander, Mana, Neeraj Naduvath, Gordon, David, Andert, Jakob · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Deep learning
Machine learning that uses many layers of simple calculations; the basis of most current AI.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 804b03bd-ad97-4475-8a79-5d2e26de3ef4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T08:01:53.686Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8fe3a88010a08932b0894b4ee8dc6ae706ceab78b42c4d9da61bb2d09b4bb65d. Source text hash: 5eae8eccc560594bac286082f442505d471490cf117bc64d7913140591584189.
Revision and source-check history
Version 1 · 2026-11-14T08:01:53.686Z
arXiv eess.SY authors report faster cloned control for a hydrogen-diesel engine, with limits beyond training range
Alexander Winkler, Neeraj Naduvath Mana, David Gordon and Jakob Andert report training a feedforward with behaviour cloning (BC)—learning to imitate an expert controller—to copy Model Predictive Control (MPC) for a hydrogen-diesel dual-fuel engine. MPC adjusts controls through online optimisation, but the authors say that computation can limit its use on low-cost embedded hardware. Their controller learned from 86,000 engine cycles recorded on a modified Cummins engine, then was experimentally tested on unseen fast-transient load steps. The authors report that both versions tracked those steps with normalised root mean square error (NRMSE) values close to the MPC expert’s, while mean NOx and particulate matter emissions were no higher than the expert’s. On a Raspberry Pi 400, inference took 2 ms or less, compared with up to 7 ms for the expert; an open-loop profile on an ESP32 also fit within the stated cycle window. These results suggest a way to reduce control computation in this engine setup, but do not establish performance across other engines or operating conditions. The authors also report a crucial boundary: beyond the training range, the cloned policy saturates its controls and exceeds the pressure-rise-rate limit.
Original source
eess.SY · Article
Source details
Original source date: Sep 27, 2026 (2026-09-27).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T08:01:53.686Z. First observed: 2026-09-29T06:09:03.131Z. Retrieved: 2026-10-29T19:03:22.641Z.
Limits
The reported training data came from one modified Cummins QSB 4.5-litre hydrogen dual-fuel engine at 1500 min-1; the evidence does not establish performance on other engines. The authors report that the cloned policy exceeds the pressure-rise-rate limit beyond the training range. The ESP32 result is described as open-loop profiling, not an experimental closed-loop test on that microcontroller.
Open question
How wide an operating range can the cloned policy handle before it exceeds the pressure-rise-rate limit?
Source-stated
The authors trained a feedforward deep neural network by behaviour cloning to imitate an MPC controller, using 86,000 engine cycles recorded on a modified hydrogen dual-fuel engine.
This paper trains a feedforward deep neural network (DNN) by behaviour cloning (BC) to imitate an MPC expert, using 86,000 engine cycles of demonstration data collected at 1500 min-1 on a modified Cummins QSB 4.5-litre hydrogen dual-fuel engine.
Source-stated
In experimental tests on unseen fast-transient load steps, both cloned variants had NRMSE values close to the MPC expert's, and mean NOx and particulate matter emissions were at or below the expert's.
Both track unseen fast-transient load steps (3-8 bar indicated mean effective pressure, IMEP) with normalised root mean square error (NRMSE) values of 7.80% and 9.03% against the expert's 8.01%, while keeping mean NOx and particulate matter emissions at or below those of the expert.
Source-stated
The authors report faster inference than the MPC expert on a Raspberry Pi 400 and an open-loop inference time within the stated cycle window on an ESP32.
Inference takes 2 ms or less on a Raspberry Pi 400 (ARM Cortex-A72 at 2.2 GHz), including 1 ms communication latency, compared to up to 7 ms for the MPC expert, a 3.5x speedup.
Source-stated
Beyond the training range, the cloned policy saturates its controls and exceeds the pressure-rise-rate limit.
Beyond the training range the cloned policy saturates its controls but exceeds the pressure-rise-rate limit.
Source and contributor context
Winkler, Alexander, Mana, Neeraj Naduvath, Gordon, David, Andert, Jakob · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Deep learning
Machine learning that uses many layers of simple calculations; the basis of most current AI.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 804b03bd-ad97-4475-8a79-5d2e26de3ef4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-14T08:01:53.686Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 8fe3a88010a08932b0894b4ee8dc6ae706ceab78b42c4d9da61bb2d09b4bb65d. Source text hash: 5eae8eccc560594bac286082f442505d471490cf117bc64d7913140591584189.
passed · 2026-11-14T08:01:53.686Z · anthropic/claude-opus-5.5
Publication: 804b03bd-ad97-4475-8a79-5d2e26de3ef4
attribution — pass: The four named authors match the arXiv listing. Results, the hardware-limitation rationale and the boundary finding are framed as author reports. Calling them 'arXiv eess.SY authors' in the title identifies the venue, not an endorsement.
Authors:Alexander Winkler, Neeraj Naduvath Mana, David Gordon, Jakob Andert
summary — pass: A general reader can follow the actors, the behaviour cloning and MPC concepts, the 86,000-cycle training data, the experimental results and the key limit. Minor gap: 'both versions' appears before the two variants (with and without process feedback) are introduced. Smallest repair: name them. The summary also omits the source's decarbonisation framing and its first-of-its-kind claim, but neither omission prevents understanding. Terms such as NRMSE, inference and 'saturates its controls' could use tooltips.
Two variants, one with process feedback and one without, are validated experimentally.
date — pass: originalPublished 2026-09-27 matches the v1 submission date. The v2 revision (27 Oct) and the feed date (28 Oct) are correctly not used.
[Submitted on 27 Sep 2026 (v1), last revised 27 Oct 2026 (this version, v2)]
support — pass: The numeric and comparative claims match the abstract. 'Close to' the expert is fair for 7.80% and 9.03% versus 8.01%. Emissions 'no higher' matches 'at or below'. The Raspberry Pi timing (2 ms or less versus up to 7 ms) and the ESP32 open-loop result are stated accurately. The inference about reduced computation is limited to this engine setup and explicitly disclaims generalisation. The question follows from the reported limitation.
Open-loop profiling on a low-cost ESP32 microcontroller at 180 MHz gives 4.3 ms per inference, 4x faster than required for the 18 ms cycle window.
media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. No better media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Machine learning that uses many layers of simple calculations; the basis of most current AI.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivMachine learning that uses many layers of simple calculations; the basis of most current AI.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv