← Paper Digest

Recoverability, Not Invariance: Certified Safety for Robots at the Limits of Performance

Thomas A. Berrueta, Joshua D. Ibrahim, Gilchrist X. Johnson, John Lathrop, Soon-Jo Chung

November 2026

TL;DR Instead of certifying an aggressive controller, certify that the robot can always hand over to a trusted recovery mode, and schedule that handover ahead of time. In simulated quadrotor flight this reaches the goal 22.5% faster than a backup-CBF filter without collisions, and a full-scale IndyNXT racecar driving in the rain made eight certified recoveries in a 30-minute run.

What they do

Safety filters such as control barrier functions (CBFs) certify the deployed controller itself. That needs a trusted model exactly where aggressive behavior operates, such as at tire saturation in the rain, and it steers the robot away from the edge of its envelope. This paper instead certifies that the robot can always hand control to a trusted recovery mode and converge back to a reference. The deployed controller, learned or aggressive, runs unfiltered until a certified switch.

How it works

  • Recovery tube. A contraction metric for the recovery mode defines a tube around its reference. Switching from anywhere inside the tube provably gives exponential convergence, also under disturbances that fit within the contraction rate times the tube radius.
  • When to switch. The contraction insertion gradient, the first-order effect of inserting the recovery mode on the certificate, drives a switching-time program over the interval before the predicted state first leaves the tube. Any feasible switching time is safe; optimality only affects performance.
  • Growing the tube. Sequential convex programming bends a closed-form tube with volume-preserving shears, and an exact boundary check accepts only steps that keep it certified, so the certified area never shrinks.
  • On the racecar, the certificate is posed on the transverse error and its rate, and evaluated algebraically from the 5 s predictions that both planners already publish.

Results

  • Simulated quadrotor, 20 trials past an obstacle of uncertain size: a standard CBF collides in all 20. Backup CBF and recoverability both reach the goal 20/20, but recoverability is 22.5% faster (5.21 s vs 6.72 s) on a 7.9% shorter path.
  • Planar benchmark: the synthesized tube covers 82.9% of the exact viability kernel and is 31.7% larger than the backup-CBF set at a 1.25 s horizon.
  • IndyNXT racecar in the rain at up to 20 m/s: eight certified recoveries in a 30-minute run. Total transverse error was 0.25-0.80 before switching and below 0.09 after recovery.

Limitations

The switch is one way; returning to the aggressive mode would need a certified re-entry condition. The guarantee is only as strong as the recovery mode’s contraction certificate. On hardware, exact evaluations are replaced by first-order approximations and filtering, and the racecar result comes from a single 30-minute run.

Videos

Full video (YouTube)

Figures

From Figure 1: The recoverability architecture. The deployed mode runs unmodified, the recovery mode is the trusted b...
From Figure 1: The recoverability architecture. The deployed mode runs unmodified, the recovery mode is the trusted behavior whose recovery tube is certified offline, and the recoverability block uses the predictions of both modes to schedule the handoff.
From Figure 2: Three deployment strategies. Left, the aggressive mode runs with no intervention and eventually leaves...
From Figure 2: Three deployment strategies. Left, the aggressive mode runs with no intervention and eventually leaves the safe set. Center, a backup-CBF filter keeps the state safe but pulls the trajectory away from the intended one whenever its constraint is active. Right, recoverability executes the intended trajectory untouched until one certified switch, after which the recovery mode converges exponentially to its reference.
From Figure 3: Time-optimal quadrotor flight past an obstacle of uncertain size. (a) The CBF collides, the backup CBF...
From Figure 3: Time-optimal quadrotor flight past an obstacle of uncertain size. (a) The CBF collides, the backup CBF detours wide, and recoverability flies the fast plan until its single certified switch. (b) Normalized goal error over time, with crosses at collisions and circles at arrivals. (c) Tube energy of one trial before (orange) and after (green) the switch, staying below the tube boundary.
From Figure 5: An IndyNXT racecar at the Putnam Park Road Course in the rain. (a) The course, the aggressive and reco...
From Figure 5: An IndyNXT racecar at the Putnam Park Road Course in the rain. (a) The course, the aggressive and recovery predictions through a corner, and the car. (b) One switching decision: the monitored energy crosses its threshold 1.68 s ahead, and the switch is scheduled at 0.75 s. (c) The executed recovery: transverse error and its rate peak near the switch, then decay. (d) Total error before and after each of the eight recoveries.
Abstract

Safety certification in robotics typically rests on forward invariance of the deployed controller, a property that demands high-fidelity models at the edge of the operating envelope, exactly where models are least reliable and robots perform best. As a result, many of the behaviors we most want robots to run, from learned policies to aggressive maneuvers at the limits of traction, cannot be certified at all. We introduce recoverability, a contraction-theoretic safety concept that resolves this tension by certifying a robot's ability to reengage a trusted recovery mode rather than the invariance of a deployed behavior. We prove that switching to the recovery mode from anywhere inside a certified tube guarantees exponential convergence, even under disturbances and model error. To decide when to switch, we derive the contraction insertion gradient, the first-order effect of a recovery insertion on the certificate, and embed it in a switching-time program in which feasibility alone guarantees safety and optimality only shapes performance. To supply the certificates, we develop sequential convex synthesis that monotonically grows the certified tube toward the exact viability kernel. We validate our approach in simulated quadrotor flight through uncertain obstacles, where it outperforms invariance-based safety filters, and in hardware deployments aboard a full-scale IndyNXT autonomous racecar, where it repeatedly recovers the car from an aggressive raceline at the limits of handling in the rain. Recoverability provides a certified exit strategy, extending formal safety guarantees to aggressive, learned, and experimental behaviors without requiring their invariance.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.