Vamsi, Dasari Raj, Zinage, Vrushabh, Bakolas, Efstathios, Chandra, Rohan · Nov 2, 2026
arXiv eess.SY: Neural Integral Control Barrier Functions for Agile Collision Avoidance
Markdown · JSON · Original source
The authors present a method to learn dynamics, a nominal controller and a barrier function from data, then combine them to build an integral control barrier function (ICBF) filter. They say control barrier functions (CBFs) help keep a planner within a safe set, but rely on assumptions their method is designed to remove. The authors train the barrier on viability: a state is safe only when an emergency maneuver can still avoid a constraint within actuator limits. They say geometric labels miss this distinction, which grows with speed. In the online program, actuator limits remain hard bounds while learned barrier rows may be relaxed. The authors also bound how far the returned input can drift from the input given by the true models. In agile-racing experiments with a 6-D dynamic bicycle, they report removing all 12 unsafe episodes across 200 held-out layouts without changing success rate, while running above 100 Hz on one core. They also report holding actuator constraints on every episode in a where a jointly trained baseline did so on 53.38%, with the advantage persisting under large errors in plant friction. These are the paper’s reported evaluations; the abstract does not establish performance beyond them.
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: Nov 2, 2026 (2026-11-02).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T08:37:54.464Z. First observed: 2026-11-03T08:51:23.977Z. Retrieved: 2026-11-03T22:31:00.694Z.
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 performance evidence in the abstract concerns agile racing with a 6-D dynamic bicycle and a ; it does not establish performance in other settings.
Open question
How does the filter perform beyond the agile-racing evaluations described in the abstract?
Source-stated
The authors propose learning dynamics, a nominal controller and a barrier function from data, and using them to build the ICBF filter constraint analytically.
we present an approach to learn all three from data, using neural ordinary differential equations (ODEs) for the dynamics, imitation for the nominal controller, and supervised regression for the barrier function.
Source-stated
The authors train the barrier on viability, counting a state as safe only if an emergency maneuver can still avoid the constraint under actuator limits.
the barrier is trained on viability rather than geometry, meaning that a state counts as safe only if some emergency maneuver can still avoid the constraint under the actuator limits
Source-stated
The authors report removing 12 unsafe episodes across 200 held-out layouts at unchanged success rate and above 100 Hz on one CPU core in agile-racing tests.
On agile racing with a 6-D dynamic bicycle, the filter removes all 12 unsafe episodes from 200 held out layouts at an unchanged success rate and above 100 Hz on one CPU core
Source-stated
In a benchmark, the authors report that the filter holds the input constraint on every episode, compared with 53.38% for a jointly trained baseline, and retains its advantage under large plant-friction errors.
holds the input constraint on every episode of a benchmark where a jointly trained baseline holds it on 53.38%, and keeps that advantage under large errors in the plant friction.
Source and contributor context
Vamsi, Dasari Raj, Zinage, Vrushabh, Bakolas, Efstathios, Chandra, Rohan · 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.
CPU
The central processing unit: a general-purpose processor that carries out a computer's instructions.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
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: 680658ca-9ad4-40f4-82a8-2b6bb52b3ebf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T08:37:54.464Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2561db5e27a135fc52b8a981d98b7d188c4bbd1225a92da0cce762f9f6124d33. Source text hash: e320e0efebf9ba2b2b1907d204895702a55a8051a8855c95c21e1f4764d05059.
Revision and source-check history
Version 1 · 2026-11-10T08:37:54.464Z
arXiv eess.SY: Neural Integral Control Barrier Functions for Agile Collision Avoidance
The authors present a method to learn dynamics, a nominal controller and a barrier function from data, then combine them to build an integral control barrier function (ICBF) filter. They say control barrier functions (CBFs) help keep a planner within a safe set, but rely on assumptions their method is designed to remove. The authors train the barrier on viability: a state is safe only when an emergency maneuver can still avoid a constraint within actuator limits. They say geometric labels miss this distinction, which grows with speed. In the online program, actuator limits remain hard bounds while learned barrier rows may be relaxed. The authors also bound how far the returned input can drift from the input given by the true models. In agile-racing experiments with a 6-D dynamic bicycle, they report removing all 12 unsafe episodes across 200 held-out layouts without changing success rate, while running above 100 Hz on one core. They also report holding actuator constraints on every episode in a where a jointly trained baseline did so on 53.38%, with the advantage persisting under large errors in plant friction. These are the paper’s reported evaluations; the abstract does not establish performance beyond them.
Original source
eess.SY · Article
Source details
Original source date: Nov 2, 2026 (2026-11-02).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T08:37:54.464Z. First observed: 2026-11-03T08:51:23.977Z. Retrieved: 2026-11-03T22:31:00.694Z.
Limits
The reported performance evidence in the abstract concerns agile racing with a 6-D dynamic bicycle and a ; it does not establish performance in other settings.
Open question
How does the filter perform beyond the agile-racing evaluations described in the abstract?
Source-stated
The authors propose learning dynamics, a nominal controller and a barrier function from data, and using them to build the ICBF filter constraint analytically.
we present an approach to learn all three from data, using neural ordinary differential equations (ODEs) for the dynamics, imitation for the nominal controller, and supervised regression for the barrier function.
Source-stated
The authors train the barrier on viability, counting a state as safe only if an emergency maneuver can still avoid the constraint under actuator limits.
the barrier is trained on viability rather than geometry, meaning that a state counts as safe only if some emergency maneuver can still avoid the constraint under the actuator limits
Source-stated
The authors report removing 12 unsafe episodes across 200 held-out layouts at unchanged success rate and above 100 Hz on one CPU core in agile-racing tests.
On agile racing with a 6-D dynamic bicycle, the filter removes all 12 unsafe episodes from 200 held out layouts at an unchanged success rate and above 100 Hz on one CPU core
Source-stated
In a benchmark, the authors report that the filter holds the input constraint on every episode, compared with 53.38% for a jointly trained baseline, and retains its advantage under large plant-friction errors.
holds the input constraint on every episode of a benchmark where a jointly trained baseline holds it on 53.38%, and keeps that advantage under large errors in the plant friction.
Source and contributor context
Vamsi, Dasari Raj, Zinage, Vrushabh, Bakolas, Efstathios, Chandra, Rohan · 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.
CPU
The central processing unit: a general-purpose processor that carries out a computer's instructions.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
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: 680658ca-9ad4-40f4-82a8-2b6bb52b3ebf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T08:37:54.464Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2561db5e27a135fc52b8a981d98b7d188c4bbd1225a92da0cce762f9f6124d33. Source text hash: e320e0efebf9ba2b2b1907d204895702a55a8051a8855c95c21e1f4764d05059.
uncertain · 2026-11-10T08:37:54.464Z · anthropic/claude-opus-5.5
Publication: 680658ca-9ad4-40f4-82a8-2b6bb52b3ebf
attribution — pass: Claims are consistently attributed to the paper's authors (the listed four-author arXiv preprint), and results are framed as reported by them. The title prefix uses the feed label 'arXiv eess.SY', while the paper's primary listing is cs.RO. eess.SY is a listed cross-list, so this is not a misattribution.
Authors:Dasari Raj Vamsi, Vrushabh Zinage, Efstathios Bakolas, Rohan Chandra
summary — uncertain: The summary is faithful but hard to follow without specialist background. Central terms ('nominal controller', 'barrier function', 'safe set', 'online program', 'barrier rows', '6-D dynamic bicycle', 'plant friction') are not explained in plain language or flagged for tooltips. It also never says why this matters, such as a safety filter for fast vehicles. It drops the stated reason for relaxing barrier rows. Smallest repair: add that relaxing them 'keeps the program solvable at every state and the actuator limits exact'. It also changes the source's 'input constraint' to 'actuator constraints'; the exact term should be kept.
which keeps the program solvable at every state and the actuator limits exact
date — pass: originalPublished 2026-11-02 matches the arXiv v1 submission date. The 2026-11-03 feed date is correctly not used.
[v1] Mon, 2 Nov 2026 06:53:14 UTC (117 KB)
support — pass: Each claim matches its abstract excerpt, and the figures (12 unsafe episodes, 200 held-out layouts, above 100 Hz, 53.38%) are reproduced accurately. The limitation and question stay within the evidence and do not inflate benchmark results into real-world reliability.
the filter removes all 12 unsafe episodes from 200 held out layouts at an unchanged success rate and above 100 Hz on one CPU core
media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
The central processing unit: a general-purpose processor that carries out a computer's instructions.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceThe central processing unit: a general-purpose processor that carries out a computer's instructions.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference