Mykel J. Kochenderfer

Research profile

Research summary

Mykel J. Kochenderfer is a researcher at Stanford University whose Evolve-reviewed snapshot spans two tightly connected lines of inquiry: formal verification of learning-enabled systems and aerospace autonomy for space situational awareness. Two 2026 arXiv papers anchor the profile. In the first, on which Kochenderfer is a middle author, RAIL/CLIPPER combines dynamics envelopes with branch-and-bound over linear bound propagation (LiRPA) to verify nonlinear neural feedback systems against reach-avoid specifications; the evaluation solves 12 instances in an assessment built around AINNCS 2025 examples and surpasses compositional solvers, though CORA remains tighter on 6 of the 12. In the second, on which Kochenderfer is the last author, the spacecraft wait-versus-maneuver decision under conjunction uncertainty is cast as chance-constrained belief-space planning solved with Monte Carlo tree search, showing promising safe deferral across 8 CARA events while acknowledging a small sample and idealized tracking. Reproducibility is the strongest scored dimension: the mean readiness across 2 scored papers is 78, with the verification work rated high and the planning study medium. The collaboration dimension scores 32 from 4 distinct coauthors in this snapshot, with recurring collaborators including Grace Ra Kim, Duncan Eddy, I. Samuel Akinwande, and Clark Barrett. Output, impact, and activity metrics are not yet available and are left unscored.

Research keywords

Highlights

• RAIL/CLIPPER: unified closed-loop verification via branch-and-bound

  On a 2026 arXiv paper (middle-author position), RAIL/CLIPPER combines dynamics envelopes with branch-and-bound over linear bound propagation (LiRPA) to verify nonlinear neural feedback systems against reach-avoid specifications. The evaluation solves 12 instances in an assessment built around AINNCS 2025 examples and surpasses compositional solvers, though CORA remains tighter on 6 of the 12. Reproduction readiness is rated high.

• Chance-constrained belief-space planning for conjunction decisions

  A second 2026 arXiv paper (last-author position) frames the spacecraft wait-versus-maneuver choice under conjunction uncertainty as chance-constrained belief-space planning solved with Monte Carlo tree search, demonstrating promising safe deferral across 8 CARA events. The review flags the small sample and idealized tracking, and rates reproduction readiness medium.

• High reproducibility across a two-field 2026 output

  Across the 2 scored papers the mean reproduction readiness is 78, the strongest scored dimension in this profile: the verification paper rates high on artifact readiness and the collision-avoidance study rates medium. This pairs with a field distribution split exactly 0.5/0.5 between conjunction assessment and neural-network verification.

Timeline

1. 2026Posts 'Closing the Loop: Branch-and-Bound for Scalable Verification of Nonlinear Neural Feedback Systems' to arXiv, introducing the RAIL/CLIPPER unified closed-loop verification framework; reproduction readiness rated high.
2. 2026Publishes 'Chance-Constrained Belief-Space Maneuver Planning for Autonomous Collision Avoidance Under Uncertainty' on arXiv, applying belief-space Monte Carlo tree search to conjunction wait-versus-maneuver decisions across 8 CARA events; reproduction readiness rated medium.
3. 2026The two cataloged papers split the field distribution evenly (0.5/0.5) between conjunction assessment and neural-network verification, with 4 distinct coauthors and a reproducibility mean of 78 defining this snapshot.

Papers in the Evolve catalog

• 2026-09-14

  Closing the Loop: Branch-and-Bound for Scalable Verification of Nonlinear Neural Feedback Systems

• 2026-09-11

  Chance-Constrained Belief-Space Maneuver Planning for Autonomous Collision Avoidance Under Uncertainty

Frequent co-authors