LIRRA  1.0.0
Loading...
Searching...
No Matches
Introduction

LIRRA:
Learning-augmented Interval-STL for Recursively-feasible Runtime Assurance
From:
Skip the Solve: Learned Trajectory Proposals for Recursively Feasible Interval-STL Runtime Assurance
By:
Evanns Morales-Cuadrado Xinpei Ni Glen Chou Samuel Coogan
Georgia Institute of Technology
Submitted to:
9th Annual Learning for Dynamics and Control Conference (L4DC 2027)

What LIRRA Does

LIRRA reduces how often a vehicle has to solve an expensive signal temporal logic (STL) synthesis problem online, while ensuring that the specification stays recursively feasible for receding-horizon planning.

Showreel of the 3-D experiments (6 min): the Skydio X2 quadrotor and the GT-MAB blimp patrolling around the restricted cube S under the four cube specifications θ, φ, θ⁺ and φ⁺, each introduced by a title card. Status bar: whether the latest proposal passed the check and, if not, which condition it failed. Left: 3-D view of S, the six waypoints and the plan in force (blue: accepted proposal, yellow: solver, grey: cached). Right: MuJoCo view, realized robustness, and the interval robustness of each proposal against the threshold δ. Every vehicle and specification has its own clip on the Videos page.

Why Recursive Feasibility

Robots are asked to meet deadlines, time windows and recurring obligations, not just to stay inside a safe set: leave a region within a few seconds of entering it, return to a station every lap, stay out of an area once you have left it. STL expresses such requirements, and for linear models with affine predicates a satisfying trajectory can be synthesized as a mixed-integer linear program (MILP).

Finding a satisfying trajectory over the planning horizon is not enough. A receding-horizon controller replans at every step, and a plan that satisfies every window it can see may still lead somewhere from which no satisfying continuation exists. Many receding-horizon STL controllers simply assume that the next problem will be feasible, or relax the specification when it is not. Recursive feasibility removes that assumption: if the problem is feasible now, a feasible continuation is guaranteed to exist at every later step, so the specification holds for the whole mission.

Uncertainty in the System and in the Specification

Satisfying a specification for a nominal trajectory says little once disturbances and modeling errors act on the real vehicle. Uncertainty can also sit in the specification itself, for example when the position or size of a restricted region is only known within bounds. Interval STL (I-STL) handles both by evaluating robustness over interval-valued signals and predicates, giving a lower bound that holds for every admissible realization. This supports requirements ranging from avoiding an uncertain obstacle to entering and later leaving a region whose boundary is uncertain.

LIRRA builds on the I-STL runtime assurance of Baird, Morales-Cuadrado and Coogan, which combines I-STL with persistently feasible terminal sets to guarantee recursive feasibility under bounded uncertainty, and keeps a safe backup input available even when the solver misses a deadline. Its price is one MILP per planning step, a heavy load for a small vehicle's onboard computer.

LIRRA's acceptance test is the I-STL synthesis program itself, so it inherits this treatment of uncertainty unchanged. In the case studies the uncertainty comes from the mismatch between each MuJoCo vehicle and its linear planning model, bounded by a tracking tube, and from bounded disturbances on the double integrator.

How LIRRA Works

A learned controller is cheap to evaluate, but on its own it guarantees neither specification satisfaction nor recursive feasibility. LIRRA keeps the synthesis problem as the definition of what is safe to fly, and uses a learned policy only to propose solutions to it:

  1. Propose. At every planning step a learned policy outputs a complete input sequence for the planning model, i.e. a candidate trajectory segment.
  2. Check. The segment is accepted only if it is a feasible point of the I-STL synthesis program: tube-tightened state and input bounds, the STL margins on every receding window, and a terminal block inside a verified control-invariant set. Checking a given trajectory is a rollout and a few comparisons, not an optimization.
  3. Fall back. If the check rejects the proposal, the vehicle keeps flying its last verified plan while the full program is solved from the state the plan predicts once the solver's time budget has elapsed (delay compensation). The result is installed when it returns and passes the same check.

Because the cached plan, shifted and extended by the terminal controller, is always a feasible point of the next program, a certified input exists at every step, including while a solve is pending or after it misses its deadline. Specification satisfaction therefore holds for any proposer; training only decides how often synthesis is needed.

What Is New

Relative to synthesis-based I-STL runtime assurance, LIRRA contributes:

  • Verified learned proposals. Whole-trajectory proposals are flown only if they satisfy the original synthesis constraints, so most planning steps need no optimization at all.
  • Delay-compensated fallback. Synthesis starts from the state the vehicle will reach when the solver's budget expires, with error-age bookkeeping for every predicted sample, so a solution is valid when it arrives. Every solver result is re-checked before it replaces the cached plan.
  • Verified terminal sets. Control-invariant terminal sets are computed and checked for invariance, with a terminal tail length derived from the formula, so that every specification window still open at the end of the horizon is satisfied by the continuation.
  • A training pipeline for the proposer. Behavior cloning of synthesis solutions, DAgger and PPO, with a differentiable feasibility margin whose sign matches the acceptance test.

Key Results

Eight specifications (the NAHS formulas θ and φ, on the planar square and on a 3-D cube, with exact and with uncertain boundaries), two MuJoCo vehicles, 30 episodes each; see Case Studies:

  1. No violations. All 472 LIRRA episodes on the Skydio X2 and the blimp satisfied their specification, including every uncertain-boundary episode scored against its drawn true region.
  2. Far fewer solves. LIRRA needs fallback synthesis on 0.3–1.9% of X2 planning steps and 4–8% of blimp steps. Mean computation per step falls 50–95× on the X2 (e.g. 141 ms to 1.7 ms) and 5–9× on the blimp.
  3. Harder specifications cost little. Fallback rates on the cube and with uncertain boundaries stay close to those on the original square, although on the blimp synthesis itself becomes infeasible from some cube initial states.
  4. Each safety mechanism is needed. Unchecked, the learned proposers violated θ or φ (cube) in 57 of 238 episodes; planning against nominal instead of interval boundaries entered the true region in up to two thirds of episodes.
  5. Mission progress is mixed on the blimp. The X2 keeps its six-waypoint pace; the blimp reaches fewer waypoints than always-on synthesis on most specifications and more on two.

Features

  • Receding-horizon I-STL synthesis (Gurobi MILP through an stlpy-style interface) with error-age-indexed tubes and verified control-invariant terminal sets
  • A JAX feasibility check whose almost-everywhere differentiable margin has exactly the sign of the acceptance test, used both at runtime and as a training signal
  • Delay-compensated fallback with cached-plan execution and an enforced solver time budget
  • Proposer training by behavior cloning of synthesis solutions, DAgger, and PPO (Brax)
  • Benchmarks: a double integrator, the MuJoCo Menagerie Skydio X2 flown through a body-rate tracker, and a MuJoCo model of Georgia Tech's Miniature Autonomous Blimp
  • Per-call timing and CPU energy (Intel RAPL) for solves, checks and policy inference

Documentation Sections