Researchers propose a safety test for robots that switch to a fallback

The authors report that a simulated drone avoided collisions in every trial. On a wet track, an autonomous IndyNXT racecar made eight certified recoveries during a 30-minute run, they say.

Numbers in brackets link to the sources listed below.

Researchers proposed a way to certify when a robot can safely switch to a trusted fallback in a preprint submitted to arXiv on 17 November 2026[1arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [1]]. They call the approach recoverability and report tests in simulated flight and on an autonomous racecar[1arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [1]].

The authors call the region in which they can certify recovery a tube[2arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [2]]. They say switching to the fallback anywhere within that region guarantees exponential convergence even if the robot encounters disturbances or its model is inaccurate[1arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [1]].

In the drone simulations, the researchers compared switching with a safety filter that blended commands from the same flight and recovery modes[2arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [2]].

At Putnam Park Road Course, the team says its IndyNXT racecar made switching decisions onboard and reached 20 m/s in the rain[2arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [2]]. The researchers report that reduced grip repeatedly brought the car close to the tube boundary in corners and that they recorded eight certified recoveries during the 30-minute run[2arXivRecoverability, Not Invariance: Certified Safety for Robots at the Limits of PerformanceOpen sourceSources [2]].

Sources

  1. 1
  2. 2

ITANU’s newsroom system produced this article with AI and checked it against the sources listed above. How we work

ITANU Magazine

AI, computing and science news