hakobyan2019risk

Risk-Aware Motion Planning and Control Using CVaR-Constrained Optimization

Authors: Hakobyan, Kim, Yang · Year: 2019 · Venue: IEEE Robotics and Automation Letters, Vol. 4 No. 4, pp. 3924–3931 · DOI: 10.1109/LRA.2019.2929980 · Raw: not yet converted — IEEE-paywalled, no arXiv preprint, and the LAN/marker box was offline at ingest (2026-07-04).

Partial ingest — built from the abstract + a verified secondary source, not the full text

The publisher PDF is paywalled, no open-access preprint exists (the SNU repository record 10371/195096 carries the metadata but no file), and marker conversion was unavailable (LAN cluster down 2026-07-04). This page is therefore synthesised from (1) the verbatim published abstract (confirmed identical via Semantic Scholar and the SNU record) and (2) the adversarially-verified extraction in deep_research_cvar_vs_chance_constrained (findings 2, 5, 8; each 3-0 verified). No equation-level transcription was possible — the Method/Equations sections below are deliberately claim-level. Convert the full text and complete them when the fleet returns.

Summary

The paper proposes a risk-aware motion planning and decision-making method that systematically trades safety against conservativeness in an environment with randomly moving obstacles. Its key ingredient is the conditional value-at-risk (CVaR), used to quantify the safety risk a robot faces. The method is two-stage: a reference trajectory is first generated with RRT*, then a receding-horizon controller limits the safety risk via CVaR constraints. The second-stage problem is a triple-level stochastic program, made computationally tractable through (1) a reformulation of the CVaR constraints, (2) a sample average approximation (SAA), and (3) a linearly constrained mixed-integer convex program. Utility is demonstrated in simulation on a 12-dimensional quadrotor model.

Key Claims

• CVaR over chance constraints, on semantics not tractability: “Unlike chance constraints, CVaR constraints are coherent, convex, and distinguish between tail events” (abstract). CVaR measures the severity of a constraint violation, whereas a chance constraint bounds only its probability.
• Two-stage architecture: a global RRT* reference trajectory (stage 1) followed by a receding-horizon controller that enforces CVaR safety constraints against the moving-obstacle uncertainty (stage 2).
• Triple-level stochastic program: the stage-2 receding-horizon problem with distance-to-safe-set losses is a nested three-level stochastic program, non-trivial to solve directly (verified secondary extraction).
• Tractable route: CVaR-constraint reformulation → sample average approximation → linearly constrained mixed-integer convex program. Critically, the integer variables arise from the nonconvex keep-out geometry, not from CVaR itself — a chance-constrained formulation of the same geometry would carry the same combinatorial burden (verified secondary extraction, deep-research finding 5).
• Validation: simulation on a 12-dimensional quadrotor model; no hardware.

Method

Claim-level only — the full text was not converted (see the partial-ingest warning).

Stage 1 — reference trajectory: a collision-aware reference path is generated with RRT* over the workspace.

Stage 2 — risk-constrained receding-horizon control: a receding-horizon controller tracks the reference while bounding the CVaR of a distance-to-safe-set (safety) loss against randomly moving obstacles. Because this couples the control decision, the sampled obstacle uncertainty, and the inner CVaR minimisation, it is a triple-level stochastic program. The published tractable reduction is: reformulate the CVaR constraint (Rockafellar–Uryasev form), replace the expectation by a sample average approximation, and encode the resulting nonconvex safe set with integer variables to obtain a linearly constrained mixed-integer convex program.

Regime: a generic robot-navigation paper — the demonstration platform is a 12-DOF quadrotor, so it is neither free-flying nor free-floating and carries no manipulator dynamics, base–arm coupling, or momentum bookkeeping. Its relevance to a free-flying space manipulator lives entirely at the risk/planning layer, not the dynamics model.

Relevance to thesis

A canonical template for the risk layer atop nominal guidance/control of the free-flying manipulator, and a direct anchor for the CVaR-vs-chance comparison behind risk-aware view scoring. Three points carry over: (i) the abstract’s coherence/convexity/tail-distinction argument is exactly the severity-matters, not tractability justification the thesis adopts for CVaR under camera-pose and covariance-misspecification uncertainty; (ii) the CVaR-reformulation → SAA → mixed-integer route is a concrete recipe for a keep-out-constrained receding-horizon controller during inspection; (iii) the verified caveat that the mixed-integer cost comes from the keep-out geometry, not from CVaR is load-bearing — it forecloses the mistaken objection that “CVaR is too expensive,” since the same geometry burdens a chance-constrained formulation identically. Gap to close: the dynamics are a quadrotor, so risk must be propagated through the 6-DOF base + redundant-arm Jacobian before this transfers, and the moving-obstacle model must be recast as camera-pose / target-tumble uncertainty for view scoring.

Connections

Topics: conditional_value_at_risk · chance_constraints · risk_aware_mpc · receding_horizon_control · motion_planning · keep_out_zone

Sources: dixit2023risk · majumdar2017how · ren2022chance · akella2024risk

Key Equations / Quotes

“The key component of this method is the conditional value-at-risk (CVaR) used to measure the safety risk that a robot faces. Unlike chance constraints, CVaR constraints are coherent, convex, and distinguish between tail events.” (Abstract)

“The second stage problem is nontrivial to solve, as it is a triple-level stochastic program. We develop a computationally tractable approach through 1) a reformulation of the CVaR constraints; 2) a sample average approximation; and 3) a linearly constrained mixed integer convex program formulation.” (Abstract)

Equation-level transcription (the Rockafellar–Uryasev CVaR reformulation, the SAA objective, the mixed-integer keep-out encoding) is deferred to full-text conversion; see conditional_value_at_risk for the canonical CVaR definition.

Open Questions

• Full-text conversion pending — transcribe the stage-2 CVaR reformulation, the SAA objective, and the mixed-integer keep-out encoding, then lift this page from developing to a complete Method section.
• How does the triple-level → mixed-integer route scale when the safety loss is propagated through a 6-DOF free-flying base + redundant-arm Jacobian rather than a quadrotor point-mass?
• Does the moving-obstacle CVaR constraint recast cleanly as a camera-pose / target-tumble uncertainty for risk-aware view scoring, or does the sensing objective break the distance-to-safe-set loss structure?
• Sample complexity of the SAA at inspection-relevant confidence levels — how many scenarios are needed for a trustworthy tail estimate, and how does that compare to a scenario-approach chance constraint at matched effective risk?