Recoverability, Not Invariance: Certified Safety for Robots at the Limits of Performance
Authors: Thomas A. Berrueta, Joshua D. Ibrahim, Gilchrist X. Johnson, John Lathrop, Soon-Jo Chung
Organizations: Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA 91125 USA · Department of Mechanical Engineering, Stanford University, Stanford, CA 94305 USA
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.
Figures & tables
Fig. 1: The recoverability architecture. Because recoverability is defined directly on closed-loop vector fields, each mode can be built from any part of an autonomy stack, including the state estimation, motion planning, and tracking control shown here, as well as policies or other learned components. The deployed mode Fnew runs unmodified, and the recovery mode Frec is the trusted behavior whose recovery tube is certified offline. The recoverability block summarizes Algorithm 2 , which uses the predictions of both modes to schedule the handoff at time τ⋆ .
Fig. 2: Safety and performance under three deployment strategies. In the left panel the aggressive deployed mode runs with no intervention, so the executed trajectory follows xnew(t) at zero performance cost but eventually leaves the safe set S with no certificate to warn of the exit. In the center panel a backup-CBF filter keeps the state inside the rollout set BT , but the filtered trajectory deviates from xnew(t) whenever the constraint is active, so the filter incurs a performance cost whenever the constraint is active. In the right panel our framework monitors the synthesized recovery tube Tsyn⊆S and executes xnew(t) untouched until the certified switch at τ⋆ , placed ahead of the latest certified time τexit , after which the recovery mode Frec contracts to the reference xr(t) exponentially in dM , so the performance cost is zero until τ⋆ and safety is guaranteed throughout.
Fig. 3: Time-optimal quadrotor flight past an obstacle of uncertain size. (a) Ten trials on one side of the obstacle, whose true shape (gray) lies inside the known envelope (dashed). The CBF collides, the backup CBF detours wide, and our framework flies the fast plan until its single certified switch. (b) Normalized goal error, the excess of position and speed over the goal tolerances relative to its initial value, with crosses at collisions and circles at arrivals. Thin curves are trials and thick curves are means aligned by arc length in (a) and by flight duration in (b). (c) Tube energy Vq/Emax of trial 1 before (orange) and after (green) the switch, which stays below the tube boundary (dashed).
Method
Success
Time to goal (s)
Path length (m)
Obs. Clear. (m)
CBF
0/20
—
—
—
Backup CBF
20/20
6.718±0.060
13.401
0.227
Recoverability
20/20
5.206±0.045
12.338
0.146
Mean paired reduction
22.5%
7.9%
29.6%
TABLE I: Comparison of Quadrotor Results.
Fig. 4: Planar recoverability example. (a) The synthesized certificate Tsyn has greater area than the seed tube T0 while preserving a closed-form geodesic distance. The dark dashed curve is the largest centered P -ellipsoid Ec , the red curve is the deployed no-insertion trajectory, and the green curve is the recovery trajectory after the latest certified switch τexit on the tube boundary, which contracts back to xr=0 . (b) Analytic set comparison for the same planar system, including the state-constraint set S , the HJ viability kernel VHJ⊆S , the finite-horizon backup-CBF rollout set BT , and Tsyn . (c) Evolution of the set boundaries under the corresponding recovery flows, parameterized by α=e−γs . The right cross-section corresponds to s=0 , and all boundaries converge to xr=0 as s→∞ .
Fig. 5: Certified recovery mode scheduling deployed on a full-scale autonomous racecar using the architecture of Fig. 1 . (a) The vehicle is an IndyNXT racecar deployed at the Putnam Park Road Course in the rain. The inset shows the current state x(t) together with the aggressive prediction xnew(t) and the recovery prediction xrec(t) through the upcoming corner. (b) A single switching decision along the predictive horizon, with times measured from the planner update. The monitored energy Vloc first violates the effective threshold Eˉ at τV=1.68 s, and the switching function Φsur enters its tolerance band at τ⋆=0.75 s, which Algorithm 2 selects as the earliest certified candidate ahead of the violation. (c) The executed recovery. The active mode steps from Fnew to Frec at τ⋆ , and the transverse error ∣e⊥∣ and its rate ∣e˙⊥∣ peak near the switch and decay once the recovery mode engages, as Theorem 1 certifies. (d) Error statistics across eight certified recoveries from the deployment. Each post-recovery total error ∥ζ∥1 is below every pre-switch value.
A robot can predict failure and still be unable to prevent it. By the time a safety mechanism reacts, the nominal plan may already have spent the control authority that recovery requires, and fixed task priorities may block whatever response remains. Our key insight is that both aspects are decided inside the controller. Recoverability must inform actions while they are chosen rather than veto them afterward, and task objectives must be adapted as recoverability shrinks. Building on this, we present RAYA, a hybrid learned-analytic framework that places a learned finite-horizon recoverability margin inside an optimal controller with hard constraints and pairs it with a bounded learned scheduler that shifts task weights to facilitate recovery. Across 7,200 simulation episodes per controller spanning quadrotor and autonomous-vehicle benchmarks, RAYA not only improves survival rates, but also transfers the learned components zero-shot to unseen trajectories, disturbances, plant shifts, and friction layouts. We developed an embedded realization of RAYA and deployed it on-board a 35g Crazyflie quadrotor. Across 40 combined hardware flights under wind with either aerodynamic mismatch or an unmodeled 40% motor-command loss, each of three baselines fails in all trials, while RAYA completes 10/10 six-cycle missions. Project Website: https://raya-control.github.io/.
Ishaan Mahajan, Charles Chen, Frederike Dümbgen +1
College of Engineering, Carnegie Mellon University, Pittsburgh, PA, USA. · Dartmouth College, Hanover, NH, USA.
Safe navigation in heterogeneous multi-robot teams faces two coupled challenges: differences in robots' braking and turning capabilities limit how each robot can respond to an approaching neighbor, while simultaneous interactions require a single control action to satisfy multiple collision-avoidance constraints. Together, these challenges can make local safety constraints incompatible with a robot's admissible controls, leaving a decentralized safety controller without a feasible action. We propose a Capability-Aware Backup-Recoverability Filter (CA-BRF), based on the insight that navigation actions should preserve access to a certified recovery maneuver. First, we construct a backup-recoverable set in a shared Cartesian operational space by evaluating whether prescribed backup maneuvers maintain pairwise safety until the robots come to rest. Each maneuver respects the corresponding robot's kinematic and actuation constraints, making the recovery certificate capability-dependent. Second, CA-BRF filters candidate controls by selecting admissible interpolations between candidate and backup actions while coordinating the selection across simultaneous interactions. This preserves backup recoverability while retaining as much of the candidate action as possible and ensures that all pairwise recovery conditions are satisfied by a single executable control for each robot. Together, we establish recursive feasibility and safety for the discrete-time closed-loop system even when the nominal safety-constrained optimization becomes infeasible. The validity and effectiveness of the proposed CA-BRF are validated in simulations with up to 30 heterogeneous robots and experiments of robots with five different kinematic classes.
Joonkyung Kim, Yanze Zhang, Wenhao Luo +1
Department of Computer Science and Engineering, Texas A&M University · Department of Computer Science, University of Illinois Chicago
Runtime assurance (RTA) protects a safety-critical system by switching from an advanced controller to a verified safe controller when a monitored condition is violated. The standard latching rule, which trips on the first breach of the safe set and then coasts, is correct for a diverging controller but pathological for a capable online-adapting one. Such a controller is unsafe by design during a bounded recovery transient. It must excite the plant to identify the fault before it can correct it, so a latching shield trips on that transient and suppresses a controller that would have recovered. We introduce the conformal recovery-deadline certificate, a split-conformal, distribution-free, finite-sample upper bound on the adapting controller's recovery time that licenses delayed fallback with a coverage guarantee, backstopped by a verified monitor at a hard critical limit. The certified deadline discriminates capable from incapable controllers, keeping the recoverer autonomous while catching the diverger. The construction separates autonomy, governed by statistical coverage, from safety, governed by the verified backstop, as an instance of reliability-asymmetric design. We prove marginal coverage, a weighted extension that restores coverage under a known fault-distribution shift, and group-conditional Mondrian coverage. We demonstrate all three on two unrelated Simplex testbeds: a 6-DOF spacecraft attitude controller and a torque-controlled inverted pendulum. Both show the same suppression pathology and the same cure, making the certificate a domain-general mechanism rather than a single-system trick.
Alireza Shojaei
Myers-Lawson School of Construction, Virginia Tech, Blacksburg, VA 24061 USA